From f012214a98b3f3c689596953072e9ea251609839 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sat, 29 Aug 2026 22:27:06 +0800 Subject: [PATCH 01/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E5=B9=BF?= =?UTF-8?q?=E5=91=8A=E5=8C=BA=E9=97=B4=E5=8F=8D=E9=A6=88=E7=9A=84=E8=AF=81?= =?UTF-8?q?=E6=8D=AE=E5=A5=91=E7=BA=A6=E4=B8=8E=E5=88=87=E7=89=87=E6=98=A0?= =?UTF-8?q?=E5=B0=84?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第一个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 6.1、7.1 节。 - AdIntervalMapper：播放时间↔切片下标映射（durations 前缀和，半开区间 相交），以及短按场景的起点推断（DISCONTINUITY → 跨域切换 → 音频候选 → 固定窗口逐级降级）。纯计算，无新增网络请求。 - AdIntervalEvidence / SegmentFact / AudioIntervalFact / SpeechIntervalFact： 区间证据快照契约，只保留 host 与去参数 path，沿用 AdDetectionRequest 的去敏原则。 - StartOrigin：起点来源枚举，携带证据权重供后续仲裁使用。 尚未接线到播放器，无行为变化。 Verification: testMobileArm64_v8aDebugUnitTest --tests com.fongmi.android.tv.ad.feedback.* : 11 tests, 0 failures, 0 errors Task-Guard: AF1-evidence --- .../tv/ad/feedback/AdIntervalEvidence.java | 81 ++++++++++++ .../tv/ad/feedback/AdIntervalMapper.java | 119 ++++++++++++++++++ .../tv/ad/feedback/AudioIntervalFact.java | 25 ++++ .../android/tv/ad/feedback/SegmentFact.java | 37 ++++++ .../tv/ad/feedback/SpeechIntervalFact.java | 25 ++++ .../android/tv/ad/feedback/StartOrigin.java | 31 +++++ .../tv/ad/feedback/AdIntervalMapperTest.java | 118 +++++++++++++++++ 7 files changed, 436 insertions(+) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdIntervalEvidence.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdIntervalMapper.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AudioIntervalFact.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/SegmentFact.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/SpeechIntervalFact.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/StartOrigin.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/AdIntervalMapperTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdIntervalEvidence.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdIntervalEvidence.java new file mode 100644 index 0000000000..d73ff694aa --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdIntervalEvidence.java @@ -0,0 +1,81 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.List; + +/** + * 用户反馈区间的证据快照。一次采集，之后所有分类通道只读。 + * + *

去敏原则沿用 {@code AdDetectionRequest}：只保留 host 与去参数 path， + * 不含完整 URL、query、token、Cookie。见设计文档第 6.1、10 节。 + * + * @param startMs 区间起点，毫秒 + * @param endMs 区间终点，毫秒 + * @param startOrigin 起点来源 + * @param siteKey 站点标识 + * @param siteName 站点名称 + * @param vodName 剧名 + * @param flagName 线路名 + * @param episodeName 集名 + * @param playlistHost 主 playlist 的 host + * @param urlPath 播放地址去参数 path + * @param hls 是否 HLS + * @param inside 区间内切片 + * @param outside 区间外切片，作为对照组 + * @param boundedByDiscontinuity 区间首尾是否都贴着 DISCONTINUITY 边界 + * @param crossDomain 区间内切片是否整体跨域 + * @param alreadyRemovedByStructuredRuleIds 已被结构化规则删除的规则 id + * @param legacyHeuristicActive 旧启发式引擎当次是否生效 + * @param matchedExistingHosts 命中现有黑名单的 host + * @param audio 音频指纹通道事实 + * @param speech 语音通道事实 + */ +public record AdIntervalEvidence( + long startMs, long endMs, StartOrigin startOrigin, + String siteKey, String siteName, String vodName, + String flagName, String episodeName, + String playlistHost, String urlPath, boolean hls, + List inside, List outside, + boolean boundedByDiscontinuity, boolean crossDomain, + List alreadyRemovedByStructuredRuleIds, + boolean legacyHeuristicActive, + List matchedExistingHosts, + AudioIntervalFact audio, SpeechIntervalFact speech) { + + public AdIntervalEvidence { + if (startMs < 0) throw new IllegalArgumentException("startMs must not be negative"); + if (endMs <= startMs) throw new IllegalArgumentException("endMs must be greater than startMs"); + startOrigin = startOrigin == null ? StartOrigin.FALLBACK_WINDOW : startOrigin; + siteKey = siteKey == null ? "" : siteKey; + siteName = siteName == null ? "" : siteName; + vodName = vodName == null ? "" : vodName; + flagName = flagName == null ? "" : flagName; + episodeName = episodeName == null ? "" : episodeName; + playlistHost = playlistHost == null ? "" : playlistHost; + urlPath = urlPath == null ? "" : urlPath; + inside = inside == null ? List.of() : List.copyOf(inside); + outside = outside == null ? List.of() : List.copyOf(outside); + alreadyRemovedByStructuredRuleIds = alreadyRemovedByStructuredRuleIds == null + ? List.of() : List.copyOf(alreadyRemovedByStructuredRuleIds); + matchedExistingHosts = matchedExistingHosts == null ? List.of() : List.copyOf(matchedExistingHosts); + audio = audio == null ? AudioIntervalFact.unavailable() : audio; + speech = speech == null ? SpeechIntervalFact.unavailable() : speech; + } + + /** 区间时长，毫秒。 */ + public long durationMs() { + return endMs - startMs; + } + + /** 区间内是否有可用的切片证据。无切片时 HLS 通道必须弃权。 */ + public boolean hasSegmentEvidence() { + return hls && !inside.isEmpty(); + } + + /** + * 用户报告的区间是否已被结构化规则整体处理掉。 + * 为真时说明用户看到的广告来自别处，不应再新增 HLS 规则。 + */ + public boolean handledByStructuredRule() { + return !alreadyRemovedByStructuredRuleIds.isEmpty(); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdIntervalMapper.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdIntervalMapper.java new file mode 100644 index 0000000000..6d047d57ce --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdIntervalMapper.java @@ -0,0 +1,119 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.fongmi.android.tv.bean.M3u8Evidence; + +import java.util.ArrayList; +import java.util.List; + +/** + * 播放时间 ↔ HLS 切片下标的映射，以及短按场景的起点推断。 + * + *

纯计算，不做网络请求：{@code M3u8Parser.parse()} 已经产出全部所需输入。 + * 见设计文档第 4.2、7.1 节。 + */ +public final class AdIntervalMapper { + + /** 短按回溯窗口默认值，见设计文档待确认决策 1。 */ + public static final long DEFAULT_FALLBACK_WINDOW_MS = 90_000L; + + private AdIntervalMapper() { + } + + /** + * 区间起点推断结果。 + * + * @param startMs 推断出的起点 + * @param origin 起点来源 + */ + public record InferredStart(long startMs, StartOrigin origin) { + } + + /** + * 找出与 {@code [startMs, endMs)} 有交集的切片下标。 + * + *

切片 {@code i} 覆盖 {@code [cum[i], cum[i+1])}，判据是半开区间相交： + * {@code cum[i+1] > startSec && cum[i] < endSec}。恰好贴边不算相交。 + */ + public static List insideIndices(List durations, long startMs, long endMs) { + List indices = new ArrayList<>(); + if (durations == null || durations.isEmpty() || endMs <= startMs) return indices; + double startSec = startMs / 1000d; + double endSec = endMs / 1000d; + double cum = 0; + for (int i = 0; i < durations.size(); i++) { + Float value = durations.get(i); + double duration = value == null ? 0 : value; + double next = cum + duration; + if (next > startSec && cum < endSec) indices.add(i); + cum = next; + } + return indices; + } + + /** 切片 {@code index} 的起始播放位置，毫秒。 */ + public static long segmentStartMs(List durations, int index) { + if (durations == null || index <= 0) return 0L; + double cum = 0; + for (int i = 0; i < index && i < durations.size(); i++) { + Float value = durations.get(i); + if (value != null) cum += value; + } + return Math.round(cum * 1000d); + } + + /** + * 短按场景下由终点回溯推断起点。按可靠性逐级尝试： + * DISCONTINUITY 边界 → 跨域切换点 → 音频候选 → 固定窗口。 + * + * @param evidence playlist 证据，可为 null + * @param endMs 用户点击时的播放位置 + * @param audioCandidateStartMs 音频/语音通道已有候选的起点，无则传负数 + * @param fallbackWindowMs 兜底回溯窗口 + */ + public static InferredStart inferStart(M3u8Evidence evidence, long endMs, + long audioCandidateStartMs, long fallbackWindowMs) { + if (endMs <= 0) return new InferredStart(0L, StartOrigin.FALLBACK_WINDOW); + + if (evidence != null && !evidence.isEmpty()) { + List durations = evidence.getDurations(); + + long discontinuity = latestBoundaryBefore(durations, evidence.getDiscontinuities(), endMs); + if (discontinuity >= 0) return new InferredStart(discontinuity, StartOrigin.DISCONTINUITY); + + long crossDomain = latestSwitchBefore(durations, evidence.getDomainSwitches(), endMs); + if (crossDomain >= 0) return new InferredStart(crossDomain, StartOrigin.CROSS_DOMAIN); + } + + if (audioCandidateStartMs >= 0 && audioCandidateStartMs < endMs) { + return new InferredStart(audioCandidateStartMs, StartOrigin.AUDIO_CANDIDATE); + } + + long window = fallbackWindowMs > 0 ? fallbackWindowMs : DEFAULT_FALLBACK_WINDOW_MS; + return new InferredStart(Math.max(0L, endMs - window), StartOrigin.FALLBACK_WINDOW); + } + + /** 最近一个位于 {@code endMs} 之前的断点边界起始位置，没有则返回 -1。 */ + private static long latestBoundaryBefore(List durations, List discontinuities, long endMs) { + if (durations == null || discontinuities == null || discontinuities.isEmpty()) return -1L; + long best = -1L; + for (Integer index : discontinuities) { + if (index == null || index < 0 || index >= durations.size()) continue; + long position = segmentStartMs(durations, index); + if (position < endMs && position > best) best = position; + } + return best; + } + + /** 最近一个位于 {@code endMs} 之前的跨域切换点起始位置，没有则返回 -1。 */ + private static long latestSwitchBefore(List durations, List domainSwitches, long endMs) { + if (durations == null || domainSwitches == null || domainSwitches.isEmpty()) return -1L; + long best = -1L; + for (int i = 0; i < domainSwitches.size() && i < durations.size(); i++) { + Boolean switched = domainSwitches.get(i); + if (switched == null || !switched) continue; + long position = segmentStartMs(durations, i); + if (position < endMs && position > best) best = position; + } + return best; + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AudioIntervalFact.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AudioIntervalFact.java new file mode 100644 index 0000000000..23b1638e51 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AudioIntervalFact.java @@ -0,0 +1,25 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.List; + +/** + * 音频指纹通道在区间内的既有匹配结果。Phase 1 只读，不采集新指纹。 + * + * @param enabled 音频指纹功能是否已开启 + * @param matchedRuleIds 区间内命中的规则 id + * @param captureReady 本次播放是否具备采集条件（Exo、非直播、可 seek、时长已知） + */ +public record AudioIntervalFact(boolean enabled, List matchedRuleIds, boolean captureReady) { + + public AudioIntervalFact { + matchedRuleIds = matchedRuleIds == null ? List.of() : List.copyOf(matchedRuleIds); + } + + public static AudioIntervalFact unavailable() { + return new AudioIntervalFact(false, List.of(), false); + } + + public boolean hasMatch() { + return !matchedRuleIds.isEmpty(); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SegmentFact.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SegmentFact.java new file mode 100644 index 0000000000..5f01d5a813 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SegmentFact.java @@ -0,0 +1,37 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.Locale; +import java.util.Objects; + +/** + * 单个 HLS 切片的去敏事实。只保留 host 与 path，不含 query 参数与 token。 + * + * @param index 在 media playlist 中的下标，从 0 开始 + * @param host 切片主机名，解析失败时为空串 + * @param path 切片路径，已去掉 query + * @param durationSec #EXTINF 时长 + * @param discontinuityBefore 该切片之前是否有 #EXT-X-DISCONTINUITY + */ +public record SegmentFact(int index, String host, String path, + double durationSec, boolean discontinuityBefore) { + + public SegmentFact { + if (index < 0) throw new IllegalArgumentException("index must not be negative"); + host = host == null ? "" : host; + path = path == null ? "" : path; + if (durationSec < 0) durationSec = 0; + } + + /** host 是否以给定后缀结尾，用于与 playlist 域名和黑名单比对。 */ + public boolean hostEndsWith(String suffix) { + if (host.isEmpty() || suffix == null || suffix.isEmpty()) return false; + String lower = host.toLowerCase(Locale.US); + String target = suffix.toLowerCase(Locale.US); + return lower.equals(target) || lower.endsWith("." + target); + } + + /** host 是否与另一切片同域。 */ + public boolean sameHost(SegmentFact other) { + return other != null && !host.isEmpty() && Objects.equals(host, other.host); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SpeechIntervalFact.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SpeechIntervalFact.java new file mode 100644 index 0000000000..ef7b1aba86 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SpeechIntervalFact.java @@ -0,0 +1,25 @@ +package com.fongmi.android.tv.ad.feedback; + +/** + * 语音关键词通道在区间内的既有命中情况。 + * + *

刻意不记录命中的具体关键词：{@code SpeechAdSignalProvider} 的日志也遵守同一约定， + * 关键词可能包含用户自定义的敏感内容。 + * + * @param enabled 语音去广是否已开启 + * @param hitCount 区间内命中次数 + */ +public record SpeechIntervalFact(boolean enabled, int hitCount) { + + public SpeechIntervalFact { + if (hitCount < 0) hitCount = 0; + } + + public static SpeechIntervalFact unavailable() { + return new SpeechIntervalFact(false, 0); + } + + public boolean hasMatch() { + return hitCount > 0; + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/StartOrigin.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/StartOrigin.java new file mode 100644 index 0000000000..50a744b6ec --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/StartOrigin.java @@ -0,0 +1,31 @@ +package com.fongmi.android.tv.ad.feedback; + +/** + * 区间起点的来源。短按只提供一个时间点，起点按可靠性逐级推断。 + * + *

顺序即可靠性：{@link #USER_MARKED} 最强，{@link #FALLBACK_WINDOW} 最弱。 + * 仲裁时用 {@link #weight()} 折算证据权重。 + */ +public enum StartOrigin { + /** 用户在标记模式下显式框选。 */ + USER_MARKED(1.0f), + /** 落在最近一个 #EXT-X-DISCONTINUITY 边界上。 */ + DISCONTINUITY(0.95f), + /** 落在最近一次切片跨域切换点上。 */ + CROSS_DOMAIN(0.85f), + /** 复用音频/语音通道已有候选的起点。 */ + AUDIO_CANDIDATE(0.80f), + /** 无任何信号，按固定窗口回溯。 */ + FALLBACK_WINDOW(0.55f); + + private final float weight; + + StartOrigin(float weight) { + this.weight = weight; + } + + /** 证据权重，0..1。 */ + public float weight() { + return weight; + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdIntervalMapperTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdIntervalMapperTest.java new file mode 100644 index 0000000000..008a6fdb2a --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdIntervalMapperTest.java @@ -0,0 +1,118 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertTrue; + +import com.fongmi.android.tv.bean.M3u8Evidence; + +import org.junit.Test; + +import java.util.List; + +public class AdIntervalMapperTest { + + private static final List UNIFORM = List.of(10f, 10f, 10f, 10f, 10f); + + @Test + public void mapsIntervalToOverlappingSegments() { + // [15s, 35s) 与切片 1(10-20)、2(20-30)、3(30-40) 相交 + assertEquals(List.of(1, 2, 3), AdIntervalMapper.insideIndices(UNIFORM, 15_000, 35_000)); + } + + @Test + public void intervalFlushWithSegmentBoundaryExcludesNeighbours() { + // [20s, 30s) 恰好是切片 2，贴边的 1 和 3 都不算相交 + assertEquals(List.of(2), AdIntervalMapper.insideIndices(UNIFORM, 20_000, 30_000)); + } + + @Test + public void emptyOrInvalidInputYieldsNoSegments() { + assertTrue(AdIntervalMapper.insideIndices(null, 0, 10_000).isEmpty()); + assertTrue(AdIntervalMapper.insideIndices(List.of(), 0, 10_000).isEmpty()); + // 终点不晚于起点 + assertTrue(AdIntervalMapper.insideIndices(UNIFORM, 20_000, 20_000).isEmpty()); + assertTrue(AdIntervalMapper.insideIndices(UNIFORM, 30_000, 20_000).isEmpty()); + } + + @Test + public void toleratesNullDurationEntries() { + List withNull = java.util.Arrays.asList(10f, null, 10f); + // null 记为 0 时长，切片 1 退化为零宽度不与任何区间相交 + assertEquals(List.of(2), AdIntervalMapper.insideIndices(withNull, 12_000, 18_000)); + } + + @Test + public void segmentStartAccumulatesPrecedingDurations() { + assertEquals(0L, AdIntervalMapper.segmentStartMs(UNIFORM, 0)); + assertEquals(30_000L, AdIntervalMapper.segmentStartMs(UNIFORM, 3)); + // 下标越界时按现有切片求和，不抛异常 + assertEquals(50_000L, AdIntervalMapper.segmentStartMs(UNIFORM, 99)); + } + + @Test + public void prefersDiscontinuityOverCrossDomainAndWindow() { + M3u8Evidence evidence = M3u8Evidence.create( + List.of("a.ts", "b.ts", "c.ts", "d.ts", "e.ts"), + List.of(2), + UNIFORM, + List.of(false, true, false, false, false)); + + AdIntervalMapper.InferredStart start = AdIntervalMapper.inferStart(evidence, 35_000, -1, 90_000); + + assertEquals(StartOrigin.DISCONTINUITY, start.origin()); + assertEquals(20_000L, start.startMs()); + } + + @Test + public void fallsBackToCrossDomainWhenNoDiscontinuity() { + M3u8Evidence evidence = M3u8Evidence.create( + List.of("a.ts", "b.ts", "c.ts", "d.ts", "e.ts"), + List.of(), + UNIFORM, + List.of(false, false, false, true, false)); + + AdIntervalMapper.InferredStart start = AdIntervalMapper.inferStart(evidence, 45_000, -1, 90_000); + + assertEquals(StartOrigin.CROSS_DOMAIN, start.origin()); + assertEquals(30_000L, start.startMs()); + } + + @Test + public void ignoresBoundariesAtOrAfterClickPosition() { + // 断点在 30s，点击位置就是 30s：不能把起点设成终点 + M3u8Evidence evidence = M3u8Evidence.create( + List.of("a.ts", "b.ts", "c.ts", "d.ts", "e.ts"), + List.of(3), + UNIFORM, + List.of(false, false, false, false, false)); + + AdIntervalMapper.InferredStart start = AdIntervalMapper.inferStart(evidence, 30_000, -1, 90_000); + + assertEquals(StartOrigin.FALLBACK_WINDOW, start.origin()); + assertEquals(0L, start.startMs()); + } + + @Test + public void usesAudioCandidateWhenPlaylistHasNoSignal() { + AdIntervalMapper.InferredStart start = AdIntervalMapper.inferStart( + M3u8Evidence.empty(), 60_000, 42_000, 90_000); + + assertEquals(StartOrigin.AUDIO_CANDIDATE, start.origin()); + assertEquals(42_000L, start.startMs()); + } + + @Test + public void clampsFallbackWindowToZero() { + AdIntervalMapper.InferredStart start = AdIntervalMapper.inferStart(null, 30_000, -1, 90_000); + + assertEquals(StartOrigin.FALLBACK_WINDOW, start.origin()); + assertEquals(0L, start.startMs()); + } + + @Test + public void appliesDefaultWindowWhenNonPositive() { + AdIntervalMapper.InferredStart start = AdIntervalMapper.inferStart(null, 200_000, -1, 0); + + assertEquals(200_000L - AdIntervalMapper.DEFAULT_FALLBACK_WINDOW_MS, start.startMs()); + } +} From 8ce64374879628a1ce029c193033bc4074d0df95 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sat, 29 Aug 2026 22:40:03 +0800 Subject: [PATCH 02/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9AHLS=20?= =?UTF-8?q?=E5=88=87=E7=89=87=E7=BB=93=E6=9E=84=E5=BD=92=E5=9B=A0=E9=80=9A?= =?UTF-8?q?=E9=81=93?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第二个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 7.1、8.1 节。 - HlsSegmentClassifier：五种信号加权判定（跨域 0.35 / 断点 0.25 / 时长离群 0.20 / 路径特征 0.15 / 头部位置 0.05），置信度低于 0.30 或无可编码信号时弃权。跨域要求区间内外形成对照，整条 playlist 同域 不算证据；时长离群要求区间内一致且与区间外众数差异超过 0.5s。 - AdAttribution：单通道归因结论，score() 按置信度 0.7 + 机制优先级 0.3 加权，ALREADY_HANDLED 强制归零。 - RemediationKind：枚举顺序即成本-风险优先级，priorityWeight() 供仲裁使用。 - AdCategory / RiskLevel：归因类别与风险等级。 位置信号不被 HlsAdRule 支持，encodableCount() 将其排除，避免生成 无法被现有引擎执行的规则。尚未接线到播放器，无行为变化。 Verification: testMobileArm64_v8aDebugUnitTest --tests com.fongmi.android.tv.ad.feedback.* : 26 tests (11 mapper + 15 classifier), 0 failures, 0 errors Task-Guard: AF2a-hls-classifier --- .../android/tv/ad/feedback/AdAttribution.java | 49 ++++ .../android/tv/ad/feedback/AdCategory.java | 26 ++ .../tv/ad/feedback/HlsSegmentClassifier.java | 193 +++++++++++++++ .../tv/ad/feedback/RemediationKind.java | 32 +++ .../android/tv/ad/feedback/RiskLevel.java | 14 ++ .../ad/feedback/HlsSegmentClassifierTest.java | 225 ++++++++++++++++++ 6 files changed, 539 insertions(+) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttribution.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdCategory.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/RemediationKind.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/RiskLevel.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttribution.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttribution.java new file mode 100644 index 0000000000..47f6e145ab --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttribution.java @@ -0,0 +1,49 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.List; + +/** + * 单个通道对区间的归因结论。 + * + * @param channelId 通道标识，"hls" / "domain" / "existing-rule" / "audio" / "speech" / "ai" + * @param category 归因类别 + * @param confidence 通道内自评置信度，0..1 + * @param risk 采纳该结论的风险等级 + * @param evidence 人类可读的证据，直接进 UI + * @param remediation 建议的去广机制 + */ +public record AdAttribution( + String channelId, AdCategory category, float confidence, + RiskLevel risk, List evidence, RemediationKind remediation) { + + public AdAttribution { + if (channelId == null || channelId.isBlank()) { + throw new IllegalArgumentException("channelId is required"); + } + category = category == null ? AdCategory.UNKNOWN : category; + confidence = clamp(confidence); + risk = risk == null ? RiskLevel.MEDIUM : risk; + evidence = evidence == null ? List.of() : List.copyOf(evidence); + remediation = remediation == null ? RemediationKind.NONE : remediation; + } + + private static float clamp(float value) { + if (Float.isNaN(value)) return 0f; + return Math.max(0f, Math.min(1f, value)); + } + + /** + * 仲裁得分：置信度占 0.7，机制的成本-风险优先级占 0.3。 + * {@link AdCategory#ALREADY_HANDLED} 强制降到最低，见设计文档第 8.2 节。 + */ + public float score() { + if (category == AdCategory.ALREADY_HANDLED) return 0f; + return confidence * 0.7f + remediation.priorityWeight() * 0.3f; + } + + /** 是否能产出可落地的规则。 */ + public boolean actionable() { + return remediation != RemediationKind.NONE + && remediation != RemediationKind.SESSION_SKIP_ONLY; + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdCategory.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdCategory.java new file mode 100644 index 0000000000..a510976414 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdCategory.java @@ -0,0 +1,26 @@ +package com.fongmi.android.tv.ad.feedback; + +/** + * 用户反馈区间的归因类别。 + * + *

见 {@code docs/ad-feedback-timerange-design.md} 第 6.2 节。 + */ +public enum AdCategory { + /** 切片来自非本站适配域名。 */ + THIRD_PARTY_CDN_SEGMENT, + /** 固定时长硬插块。 */ + FIXED_DURATION_BLOCK, + /** 被 DISCONTINUITY 包裹的独立块。 */ + DISCONTINUITY_BLOCK, + /** 命中已知广告域名。 */ + KNOWN_AD_HOST, + /** 音频频谱指纹匹配。 */ + AUDIO_FINGERPRINT, + /** 语音关键词命中。 */ + SPEECH_KEYWORD, + /** 压制进正片，无结构特征可用。 */ + IN_STREAM_BURNED_IN, + /** 已被现有规则处理，用户看到的广告来自别处。 */ + ALREADY_HANDLED, + UNKNOWN +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java new file mode 100644 index 0000000000..d17493f823 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -0,0 +1,193 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.List; +import java.util.Locale; +import java.util.Map; + +/** + * 基于 HLS 切片结构的归因通道。零新增运行时成本：全部输入来自 + * {@code M3u8Parser.parse()} 已产出的证据，不发起新的网络请求。 + * + *

五种信号的权重见设计文档第 7.1 节。生成的条件必须落在 + * {@code HlsManifestCleaner.Rule} 已支持的范围内，否则规则无法被现有引擎执行。 + */ +public final class HlsSegmentClassifier { + + public static final String CHANNEL_ID = "hls"; + + /** 置信度低于此值时弃权。 */ + static final float MIN_CONFIDENCE = 0.30f; + + static final float WEIGHT_CROSS_DOMAIN = 0.35f; + static final float WEIGHT_DISCONTINUITY = 0.25f; + static final float WEIGHT_DURATION_OUTLIER = 0.20f; + static final float WEIGHT_PATH_HINT = 0.15f; + static final float WEIGHT_HEAD_POSITION = 0.05f; + + /** 区间内切片时长视为「整齐」的标准差上限，秒。 */ + private static final double UNIFORM_STDDEV_MAX = 0.05d; + /** 与区间外众数时长的差异下限，秒。 */ + private static final double DURATION_GAP_MIN = 0.5d; + /** 落在 playlist 头部多少比例内算「片头」。 */ + private static final double HEAD_RATIO = 0.15d; + + /** 保守的广告路径特征，只匹配明确的目录段。 */ + private static final List PATH_HINTS = + List.of("/ad/", "/ads/", "/adv/", "/preroll/", "/midroll/", "/creative/"); + + private HlsSegmentClassifier() { + } + + /** 命中的信号，用于生成规则时决定 minimumSignals。 */ + public record Signals(boolean crossDomain, boolean discontinuity, boolean durationOutlier, + boolean pathHint, boolean headPosition) { + + /** 可编码进 {@code HlsAdRule} 的信号数。位置信号不被 HlsAdRule 支持，不计入。 */ + public int encodableCount() { + int count = 0; + if (crossDomain) count++; + if (discontinuity) count++; + if (durationOutlier) count++; + if (pathHint) count++; + return count; + } + + public float weightSum() { + float sum = 0f; + if (crossDomain) sum += WEIGHT_CROSS_DOMAIN; + if (discontinuity) sum += WEIGHT_DISCONTINUITY; + if (durationOutlier) sum += WEIGHT_DURATION_OUTLIER; + if (pathHint) sum += WEIGHT_PATH_HINT; + if (headPosition) sum += WEIGHT_HEAD_POSITION; + return sum; + } + } + + /** + * 对区间做结构归因。无法判断时返回 null 表示弃权。 + */ + public static AdAttribution classify(AdIntervalEvidence evidence) { + if (evidence == null || !evidence.hasSegmentEvidence()) return null; + // 区间已被结构化规则整体处理，用户看到的广告来自别处，不应再加 HLS 规则 + if (evidence.handledByStructuredRule()) return null; + + Signals signals = detect(evidence); + float confidence = signals.weightSum(); + if (confidence < MIN_CONFIDENCE) return null; + // 无法编码进规则的结论没有落地价值 + if (signals.encodableCount() == 0) return null; + + return new AdAttribution(CHANNEL_ID, categoryOf(signals), confidence, + RiskLevel.MEDIUM, describe(evidence, signals), RemediationKind.HLS_STRUCTURED_RULE); + } + + /** 逐项检测五种信号。 */ + public static Signals detect(AdIntervalEvidence evidence) { + List inside = evidence.inside(); + List outside = evidence.outside(); + return new Signals( + detectCrossDomain(inside, outside, evidence.playlistHost()), + evidence.boundedByDiscontinuity() || inside.get(0).discontinuityBefore(), + detectDurationOutlier(inside, outside), + detectPathHint(inside), + detectHeadPosition(inside, outside)); + } + + /** + * 跨域：区间内切片整体不属于 playlist 域名，且区间外切片属于。 + * 只有形成对照才算证据 —— 整条 playlist 都跨域说明这就是该站的正常结构。 + */ + private static boolean detectCrossDomain(List inside, List outside, + String playlistHost) { + if (playlistHost == null || playlistHost.isEmpty()) return false; + boolean insideForeign = inside.stream() + .allMatch(fact -> !fact.host().isEmpty() && !fact.hostEndsWith(playlistHost)); + if (!insideForeign) return false; + if (outside.isEmpty()) return false; + return outside.stream() + .anyMatch(fact -> fact.hostEndsWith(playlistHost)); + } + + /** + * 时长离群：区间内时长高度一致，且与区间外众数明显不同。 + * 单独出现时权重不足以触发规则，必须与其他信号叠加。 + */ + private static boolean detectDurationOutlier(List inside, List outside) { + if (inside.isEmpty() || outside.isEmpty()) return false; + if (stdDev(inside) > UNIFORM_STDDEV_MAX) return false; + double insideMean = mean(inside); + Double outsideMode = mode(outside); + if (outsideMode == null) return false; + return Math.abs(insideMean - outsideMode) > DURATION_GAP_MIN; + } + + private static boolean detectPathHint(List inside) { + return inside.stream().anyMatch(fact -> { + String path = fact.path().toLowerCase(Locale.US); + return PATH_HINTS.stream().anyMatch(path::contains); + }); + } + + /** 位置：区间落在 playlist 头部 15% 以内。 */ + private static boolean detectHeadPosition(List inside, List outside) { + int total = inside.size() + outside.size(); + if (total == 0) return false; + int firstIndex = inside.stream().mapToInt(SegmentFact::index).min().orElse(Integer.MAX_VALUE); + return firstIndex <= Math.max(0, (int) Math.floor(total * HEAD_RATIO)); + } + + private static AdCategory categoryOf(Signals signals) { + if (signals.crossDomain()) return AdCategory.THIRD_PARTY_CDN_SEGMENT; + if (signals.discontinuity()) return AdCategory.DISCONTINUITY_BLOCK; + if (signals.durationOutlier()) return AdCategory.FIXED_DURATION_BLOCK; + return AdCategory.UNKNOWN; + } + + private static List describe(AdIntervalEvidence evidence, Signals signals) { + List inside = evidence.inside(); + List lines = new ArrayList<>(); + lines.add(String.format(Locale.US, "区间内 %d 个切片，时长合计 %.1fs", + inside.size(), inside.stream().mapToDouble(SegmentFact::durationSec).sum())); + if (signals.crossDomain()) { + lines.add(String.format(Locale.US, "切片域名 %s 与 playlist 域名 %s 不一致", + inside.get(0).host(), evidence.playlistHost())); + } + if (signals.discontinuity()) lines.add("区间边界存在 #EXT-X-DISCONTINUITY"); + if (signals.durationOutlier()) { + lines.add(String.format(Locale.US, "区间内切片时长一致（%.2fs），与正片切片时长差异明显", + mean(inside))); + } + if (signals.pathHint()) lines.add("切片路径包含广告目录特征"); + if (signals.headPosition()) lines.add("区间位于播放列表头部"); + lines.add("起点来源：" + evidence.startOrigin()); + return lines; + } + + private static double mean(List facts) { + return facts.stream().mapToDouble(SegmentFact::durationSec).average().orElse(0d); + } + + private static double stdDev(List facts) { + if (facts.size() < 2) return 0d; + double mean = mean(facts); + double variance = facts.stream() + .mapToDouble(fact -> Math.pow(fact.durationSec() - mean, 2)) + .sum() / facts.size(); + return Math.sqrt(variance); + } + + /** 区间外切片的众数时长，按 0.1s 粒度归桶。 */ + private static Double mode(List facts) { + Map buckets = new HashMap<>(); + for (SegmentFact fact : facts) { + long bucket = Math.round(fact.durationSec() * 10); + buckets.merge(bucket, 1, Integer::sum); + } + return buckets.entrySet().stream() + .max(Map.Entry.comparingByValue()) + .map(entry -> entry.getKey() / 10d) + .orElse(null); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RemediationKind.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RemediationKind.java new file mode 100644 index 0000000000..571e6c73fb --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RemediationKind.java @@ -0,0 +1,32 @@ +package com.fongmi.android.tv.ad.feedback; + +/** + * 归因结论对应的去广机制。 + * + *

{@link #ordinal()} 即成本-风险优先级：越靠前越优先。见设计文档第 8.1 节。 + * 仲裁器依赖这个顺序，重排枚举会改变仲裁结果。 + */ +public enum RemediationKind { + /** 启用一条已存在但默认关闭的规则，运行时零新增成本。 */ + ENABLE_EXISTING_RULE, + /** 写入 {@code UserAdRule.hosts}。 */ + HOST_BLACKLIST, + /** 写入 {@code HlsAdRule}，由 HlsManifestCleaner 执行。 */ + HLS_STRUCTURED_RULE, + /** 写入 {@code UserAdRule.regex}。 */ + URL_REGEX_RULE, + /** 写入 {@code AudioFingerprintRule}（Phase 2）。 */ + AUDIO_FINGERPRINT_RULE, + /** 追加语音关键词（Phase 2，仅建议）。 */ + SPEECH_KEYWORD, + /** 无法泛化，只能本次跳过。 */ + SESSION_SKIP_ONLY, + /** 纯诊断结论，不产出规则。 */ + NONE; + + /** 成本-风险排序权重，0..1，越大越优先。 */ + public float priorityWeight() { + int span = values().length - 1; + return span == 0 ? 1f : 1f - (float) ordinal() / span; + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RiskLevel.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RiskLevel.java new file mode 100644 index 0000000000..eb02c1a87d --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RiskLevel.java @@ -0,0 +1,14 @@ +package com.fongmi.android.tv.ad.feedback; + +/** + * 规则风险等级，沿用智能去广 V2 总设计第 12 节的三级划分。 + * + *

低：广告 host 拦截、严格媒体 URL 排除； + * 中：结构化 HLS 区块删除、DOM 声明式动作； + * 高：任意 JavaScript、全文正则替换、宽泛固定时长规则。 + */ +public enum RiskLevel { + LOW, + MEDIUM, + HIGH +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java new file mode 100644 index 0000000000..7ca476c7aa --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java @@ -0,0 +1,225 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertNotNull; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +public class HlsSegmentClassifierTest { + + private static final String PLAYLIST_HOST = "v.example.com"; + + /** 构造区间内外切片：inside 用广告域名，outside 用 playlist 域名。 */ + private static AdIntervalEvidence evidence(List inside, List outside) { + return evidence(inside, outside, false, List.of()); + } + + private static AdIntervalEvidence evidence(List inside, List outside, + boolean bounded, List removedRuleIds) { + return new AdIntervalEvidence( + 10_000, 40_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + PLAYLIST_HOST, "/play/index.m3u8", true, + inside, outside, bounded, false, + removedRuleIds, false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + private static List segments(int from, int count, String host, String pathPrefix, + double duration, boolean discontinuityOnFirst) { + List facts = new ArrayList<>(); + for (int i = 0; i < count; i++) { + facts.add(new SegmentFact(from + i, host, pathPrefix + (from + i) + ".ts", + duration, discontinuityOnFirst && i == 0)); + } + return facts; + } + + /** 区间外切片分成前后两段，避免与区间内下标重叠。 */ + private static List outsideAround(int insideFrom, int insideCount, int tailCount, + String host, double duration) { + List facts = new ArrayList<>(segments(0, insideFrom, host, "/seg/", duration, false)); + facts.addAll(segments(insideFrom + insideCount, tailCount, host, "/seg/", duration, false)); + return facts; + } + + @Test + public void crossDomainWithDiscontinuityIsHighConfidence() { + // 区间在中段（下标 10-14，总 25 片），位置信号不参与 + AdIntervalEvidence evidence = evidence( + segments(10, 5, "ad-cdn.other.com", "/seg/", 6.4, true), + outsideAround(10, 5, 10, PLAYLIST_HOST, 8.0)); + + AdAttribution attribution = HlsSegmentClassifier.classify(evidence); + + assertNotNull(attribution); + assertEquals(AdCategory.THIRD_PARTY_CDN_SEGMENT, attribution.category()); + assertEquals(RemediationKind.HLS_STRUCTURED_RULE, attribution.remediation()); + // 跨域 0.35 + 断点 0.25 + 时长离群 0.20 = 0.80 + assertEquals(0.80f, attribution.confidence(), 0.001f); + assertTrue(attribution.actionable()); + assertFalse(HlsSegmentClassifier.detect(evidence).headPosition()); + } + + @Test + public void headPositionAddsSmallWeightOnTopOfStructuralSignals() { + // 同样三个结构信号，但区间在头部：额外 +0.05 + AdIntervalEvidence evidence = evidence( + segments(3, 5, "ad-cdn.other.com", "/seg/", 6.4, true), + outsideAround(3, 5, 17, PLAYLIST_HOST, 8.0)); + + AdAttribution attribution = HlsSegmentClassifier.classify(evidence); + + assertNotNull(attribution); + assertEquals(0.85f, attribution.confidence(), 0.001f); + } + + @Test + public void sameDomainThroughoutIsNotCrossDomainEvidence() { + // 整条 playlist 同域：这是该站正常结构，跨域信号必须不成立 + AdIntervalEvidence evidence = evidence( + segments(3, 5, PLAYLIST_HOST, "/seg/", 8.0, false), + segments(0, 20, PLAYLIST_HOST, "/seg/", 8.0, false)); + + HlsSegmentClassifier.Signals signals = HlsSegmentClassifier.detect(evidence); + + assertFalse(signals.crossDomain()); + // 无任何信号，整体弃权 + assertNull(HlsSegmentClassifier.classify(evidence)); + } + + @Test + public void crossDomainRequiresOutsideContrast() { + // 区间外为空，无法形成对照 + AdIntervalEvidence evidence = evidence( + segments(0, 5, "ad-cdn.other.com", "/seg/", 6.4, false), + List.of()); + + assertFalse(HlsSegmentClassifier.detect(evidence).crossDomain()); + } + + @Test + public void subdomainOfPlaylistHostCountsAsSameDomain() { + AdIntervalEvidence evidence = evidence( + segments(3, 3, "cdn." + PLAYLIST_HOST, "/seg/", 8.0, false), + segments(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + + assertFalse(HlsSegmentClassifier.detect(evidence).crossDomain()); + } + + @Test + public void durationOutlierNeedsUniformInsideAndDifferentOutsideMode() { + // 区间内时长不齐（6.4 与 9.0），不算离群证据 + List inside = new ArrayList<>(segments(3, 2, "ad.other.com", "/seg/", 6.4, false)); + inside.add(new SegmentFact(5, "ad.other.com", "/seg/5.ts", 9.0, false)); + + AdIntervalEvidence evidence = evidence(inside, segments(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + + assertFalse(HlsSegmentClassifier.detect(evidence).durationOutlier()); + } + + @Test + public void durationCloseToOutsideModeIsNotOutlier() { + // 6.4 与众数 6.7 只差 0.3s，低于 0.5s 门槛 + AdIntervalEvidence evidence = evidence( + segments(3, 4, "ad.other.com", "/seg/", 6.4, false), + segments(0, 10, PLAYLIST_HOST, "/seg/", 6.7, false)); + + assertFalse(HlsSegmentClassifier.detect(evidence).durationOutlier()); + } + + @Test + public void pathHintMatchesOnlyExplicitAdDirectories() { + AdIntervalEvidence hit = evidence( + segments(3, 2, PLAYLIST_HOST, "/ads/spot", 8.0, false), + segments(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + assertTrue(HlsSegmentClassifier.detect(hit).pathHint()); + + // "upload/" 含 ad 子串但不是目录段，不应命中 + AdIntervalEvidence miss = evidence( + segments(3, 2, PLAYLIST_HOST, "/upload/download", 8.0, false), + segments(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + assertFalse(HlsSegmentClassifier.detect(miss).pathHint()); + } + + @Test + public void headPositionOnlyForLeadingSegments() { + AdIntervalEvidence head = evidence( + segments(0, 3, "ad.other.com", "/seg/", 6.4, false), + segments(3, 20, PLAYLIST_HOST, "/seg/", 8.0, false)); + assertTrue(HlsSegmentClassifier.detect(head).headPosition()); + + AdIntervalEvidence middle = evidence( + segments(12, 3, "ad.other.com", "/seg/", 6.4, false), + segments(0, 20, PLAYLIST_HOST, "/seg/", 8.0, false)); + assertFalse(HlsSegmentClassifier.detect(middle).headPosition()); + } + + @Test + public void abstainsWhenAlreadyHandledByStructuredRule() { + AdIntervalEvidence evidence = evidence( + segments(3, 5, "ad-cdn.other.com", "/ads/", 6.4, true), + segments(0, 20, PLAYLIST_HOST, "/seg/", 8.0, false), + true, List.of("baofeng-preroll")); + + assertNull(HlsSegmentClassifier.classify(evidence)); + } + + @Test + public void abstainsWithoutSegmentEvidence() { + assertNull(HlsSegmentClassifier.classify(null)); + + AdIntervalEvidence empty = evidence(List.of(), segments(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + assertNull(HlsSegmentClassifier.classify(empty)); + } + + @Test + public void abstainsForNonHlsPlayback() { + AdIntervalEvidence nonHls = new AdIntervalEvidence( + 10_000, 40_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + PLAYLIST_HOST, "/play/video.mp4", false, + segments(3, 5, "ad.other.com", "/ads/", 6.4, true), + segments(0, 20, PLAYLIST_HOST, "/seg/", 8.0, false), + true, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + assertNull(HlsSegmentClassifier.classify(nonHls)); + } + + @Test + public void headPositionAloneIsBelowConfidenceFloor() { + // 只有位置信号 0.05，低于 0.30 门槛 + AdIntervalEvidence evidence = evidence( + segments(0, 3, PLAYLIST_HOST, "/seg/", 8.0, false), + segments(3, 20, PLAYLIST_HOST, "/seg/", 8.0, false)); + + HlsSegmentClassifier.Signals signals = HlsSegmentClassifier.detect(evidence); + assertTrue(signals.headPosition()); + assertNull(HlsSegmentClassifier.classify(evidence)); + } + + @Test + public void positionSignalIsNotEncodableIntoRule() { + HlsSegmentClassifier.Signals signals = + new HlsSegmentClassifier.Signals(false, false, false, false, true); + + // HlsAdRule 不支持位置条件，可编码信号数为 0 + assertEquals(0, signals.encodableCount()); + } + + @Test + public void discontinuityOnFirstInsideSegmentCountsAsBoundary() { + AdIntervalEvidence evidence = evidence( + segments(3, 4, "ad.other.com", "/seg/", 6.4, true), + segments(0, 20, PLAYLIST_HOST, "/seg/", 8.0, false)); + + assertTrue(HlsSegmentClassifier.detect(evidence).discontinuity()); + } +} From 39bbdf2510a82b3a368eeecd53f894c3822c4a43 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sat, 29 Aug 2026 22:50:37 +0800 Subject: [PATCH 03/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E5=9F=9F?= =?UTF-8?q?=E5=90=8D=E4=BF=A1=E8=AA=89=E4=B8=8E=E6=97=A2=E6=9C=89=E8=A7=84?= =?UTF-8?q?=E5=88=99=E5=BD=92=E5=9B=A0=E9=80=9A=E9=81=93=EF=BC=8C=E4=BB=A5?= =?UTF-8?q?=E5=8F=8A=E8=B7=A8=E9=80=9A=E9=81=93=E4=BB=B2=E8=A3=81=E5=99=A8?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第三个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 7.2、7.3、8 节。 - DomainReputationClassifier：三层比对。命中现有黑名单时输出 ALREADY_HANDLED 诊断（黑名单只在 WebView 层生效，不拦播放器直连切片）； 不在本站基线内则判 THIRD_PARTY_CDN_SEGMENT；接口学习候选命中加 0.15。 无基线数据时弃权，避免首次播放该站即误判。 - ExistingRuleClassifier：优先建议启用已存在但默认关闭的规则，这是成本 最低的修复；编译失败的规则只报诊断不建议启用。另报旧启发式引擎生效、 正片保护规则可能误保护。 - AdAttributionArbiter：同类别按概率或合并置信度并取成本最低机制，跨类别 按 score() 排序（置信度 0.7 + 机制优先级 0.3）。不可落地的结论一律降级 为诊断，只剩诊断时兜底为 SESSION_SKIP_ONLY 并合并证据供 UI 展示。 三个分类器均以 record Input 接收外部数据，不触碰 SharedPreferences， 可在纯 JVM 下单测。尚未接线到播放器，无行为变化。 Verification: testMobileArm64_v8aDebugUnitTest --tests com.fongmi.android.tv.ad.feedback.* : 49 tests (11 mapper + 15 hls + 7 domain + 7 existing + 9 arbiter), 0 failures, 0 errors Task-Guard: AF2b-domain-arbiter --- .../tv/ad/feedback/AdAttributionArbiter.java | 127 +++++++++++++++ .../feedback/DomainReputationClassifier.java | 130 ++++++++++++++++ .../ad/feedback/ExistingRuleClassifier.java | 117 ++++++++++++++ .../ad/feedback/AdAttributionArbiterTest.java | 144 ++++++++++++++++++ .../DomainReputationClassifierTest.java | 109 +++++++++++++ .../feedback/ExistingRuleClassifierTest.java | 112 ++++++++++++++ 6 files changed, 739 insertions(+) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiter.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiterTest.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifierTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiter.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiter.java new file mode 100644 index 0000000000..832ee1ae36 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiter.java @@ -0,0 +1,127 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.ArrayList; +import java.util.Comparator; +import java.util.EnumMap; +import java.util.LinkedHashSet; +import java.util.List; +import java.util.Map; +import java.util.Set; + +/** + * 跨通道仲裁：合并同类结论、按成本-风险排序、选出最优方案。 + * + *

核心取舍见设计文档第 8 节：「最好的去广告方式」不等于「置信度最高的方式」。 + * 同等置信度下优先选成本更低、风险更小的机制。 + */ +public final class AdAttributionArbiter { + + private AdAttributionArbiter() { + } + + /** + * 仲裁结果。 + * + * @param preferred 首选方案，全部弃权时为 null + * @param alternatives 备选方案，按得分降序 + * @param diagnostics 纯诊断结论，不产出规则但需展示 + */ + public record Verdict(AdAttribution preferred, List alternatives, + List diagnostics) { + + public Verdict { + alternatives = alternatives == null ? List.of() : List.copyOf(alternatives); + diagnostics = diagnostics == null ? List.of() : List.copyOf(diagnostics); + } + + /** 是否有可落地的规则方案。 */ + public boolean hasActionablePlan() { + return preferred != null && preferred.actionable(); + } + + /** 全部通道都没给出结论。 */ + public boolean empty() { + return preferred == null && diagnostics.isEmpty(); + } + } + + /** + * 仲裁若干通道结论。入参可含 null（通道弃权），会被忽略。 + */ + public static Verdict arbitrate(List attributions) { + List valid = new ArrayList<>(); + if (attributions != null) { + for (AdAttribution attribution : attributions) { + if (attribution != null) valid.add(attribution); + } + } + if (valid.isEmpty()) return new Verdict(null, List.of(), List.of()); + + List diagnostics = new ArrayList<>(); + List actionable = new ArrayList<>(); + for (AdAttribution attribution : valid) { + if (attribution.actionable()) actionable.add(attribution); + else diagnostics.add(attribution); + } + + List merged = mergeSameCategory(actionable); + merged.sort(Comparator.comparingDouble(AdAttribution::score).reversed() + .thenComparingInt(a -> a.remediation().ordinal())); + + if (merged.isEmpty()) { + // 只有诊断结论：能做的只有本次跳过 + return new Verdict(sessionSkipFallback(diagnostics), List.of(), diagnostics); + } + return new Verdict(merged.get(0), merged.subList(1, merged.size()), diagnostics); + } + + /** + * 同一 {@link AdCategory} 的多通道结论合并为一条：置信度按概率或 + * {@code 1 - Π(1 - cᵢ)}，机制取成本最低的那个，证据全量保留。 + */ + private static List mergeSameCategory(List attributions) { + Map> grouped = new EnumMap<>(AdCategory.class); + for (AdAttribution attribution : attributions) { + grouped.computeIfAbsent(attribution.category(), key -> new ArrayList<>()).add(attribution); + } + + List merged = new ArrayList<>(); + for (Map.Entry> entry : grouped.entrySet()) { + List group = entry.getValue(); + if (group.size() == 1) { + merged.add(group.get(0)); + continue; + } + float inverse = 1f; + for (AdAttribution attribution : group) inverse *= (1f - attribution.confidence()); + float confidence = 1f - inverse; + + AdAttribution cheapest = group.stream() + .min(Comparator.comparingInt(a -> a.remediation().ordinal())) + .orElse(group.get(0)); + RiskLevel risk = group.stream() + .map(AdAttribution::risk) + .min(Comparator.comparingInt(Enum::ordinal)) + .orElse(cheapest.risk()); + + Set lines = new LinkedHashSet<>(); + List channels = new ArrayList<>(); + for (AdAttribution attribution : group) { + lines.addAll(attribution.evidence()); + channels.add(attribution.channelId()); + } + merged.add(new AdAttribution(String.join("+", channels), entry.getKey(), + confidence, risk, List.copyOf(lines), cheapest.remediation())); + } + return merged; + } + + /** 只有诊断结论时的兜底：本次跳过。 */ + private static AdAttribution sessionSkipFallback(List diagnostics) { + List lines = new ArrayList<>(); + lines.add("未能归纳出可复用的去广规则"); + for (AdAttribution diagnostic : diagnostics) lines.addAll(diagnostic.evidence()); + return new AdAttribution("arbiter", AdCategory.IN_STREAM_BURNED_IN, + 0.3f, RiskLevel.LOW, lines, RemediationKind.SESSION_SKIP_ONLY); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java new file mode 100644 index 0000000000..f9fa403c5b --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java @@ -0,0 +1,130 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.ArrayList; +import java.util.LinkedHashSet; +import java.util.List; +import java.util.Locale; +import java.util.Set; + +/** + * 基于域名信誉的归因通道，实现「非当前适配的主流域名」这条判据。 + * + *

三层比对见设计文档第 7.2 节。所有外部数据由调用方以 {@link Input} 传入， + * 类本身不触碰 SharedPreferences，因此可在纯 JVM 下单测。 + */ +public final class DomainReputationClassifier { + + public static final String CHANNEL_ID = "domain"; + + /** 命中已知黑名单时的置信度。 */ + static final float CONFIDENCE_KNOWN_HOST = 0.95f; + /** 仅「不在本站基线内」时的置信度。 */ + static final float CONFIDENCE_UNKNOWN_HOST = 0.55f; + /** 接口候选也认为是广告域名时的加成。 */ + static final float BONUS_INTERFACE_CANDIDATE = 0.15f; + + private DomainReputationClassifier() { + } + + /** + * 分类输入。 + * + * @param blacklistedHosts 现有黑名单，来自 {@code RuleConfig.get().getAds()} + * @param siteBaselineHosts 本站历史正常域名，来自 SitePlaylistHostBaseline + * @param interfaceCandidateHosts 接口学习的待审候选域名 + * @param interfaceSourceName 命中候选时展示的来源接口名 + */ + public record Input(List blacklistedHosts, List siteBaselineHosts, + List interfaceCandidateHosts, String interfaceSourceName) { + + public Input { + blacklistedHosts = blacklistedHosts == null ? List.of() : List.copyOf(blacklistedHosts); + siteBaselineHosts = siteBaselineHosts == null ? List.of() : List.copyOf(siteBaselineHosts); + interfaceCandidateHosts = interfaceCandidateHosts == null ? List.of() : List.copyOf(interfaceCandidateHosts); + interfaceSourceName = interfaceSourceName == null ? "" : interfaceSourceName; + } + + public static Input empty() { + return new Input(List.of(), List.of(), List.of(), ""); + } + } + + /** 无结论时返回 null 表示弃权。 */ + public static AdAttribution classify(AdIntervalEvidence evidence, Input input) { + if (evidence == null || evidence.inside().isEmpty()) return null; + Input safe = input == null ? Input.empty() : input; + + Set foreignHosts = foreignHosts(evidence); + if (foreignHosts.isEmpty()) return null; + + List matchedBlacklist = matching(foreignHosts, safe.blacklistedHosts()); + List matchedCandidates = matching(foreignHosts, safe.interfaceCandidateHosts()); + + if (!matchedBlacklist.isEmpty()) { + // 已在黑名单里却仍被用户看到：拦截路径没覆盖播放器直连的切片请求。 + // 结论是诊断，落地方式应改用 HLS 结构化规则。 + List evidenceLines = new ArrayList<>(); + evidenceLines.add("切片域名已在广告黑名单中：" + String.join("、", matchedBlacklist)); + evidenceLines.add("黑名单主要在 WebView 请求拦截生效，不拦播放器直连的切片请求"); + evidenceLines.add("建议改用 HLS 结构化规则删除这些切片"); + return new AdAttribution(CHANNEL_ID, AdCategory.ALREADY_HANDLED, + CONFIDENCE_KNOWN_HOST, RiskLevel.LOW, evidenceLines, RemediationKind.NONE); + } + + // 无基线数据时不能断言「非本站域名」，避免首次播放即误判 + if (safe.siteBaselineHosts().isEmpty() && matchedCandidates.isEmpty()) return null; + + boolean outsideBaseline = !safe.siteBaselineHosts().isEmpty() + && foreignHosts.stream().noneMatch(host -> endsWithAny(host, safe.siteBaselineHosts())); + if (!outsideBaseline && matchedCandidates.isEmpty()) return null; + + float confidence = outsideBaseline ? CONFIDENCE_UNKNOWN_HOST : 0f; + List evidenceLines = new ArrayList<>(); + if (outsideBaseline) { + evidenceLines.add("切片域名 " + String.join("、", foreignHosts) + " 不在本站常用域名中"); + } + if (!matchedCandidates.isEmpty()) { + confidence += BONUS_INTERFACE_CANDIDATE; + String source = safe.interfaceSourceName().isEmpty() ? "接口规则" : safe.interfaceSourceName(); + evidenceLines.add(source + "也将其列为广告域名候选"); + } + evidenceLines.add("起点来源：" + evidence.startOrigin()); + + return new AdAttribution(CHANNEL_ID, AdCategory.THIRD_PARTY_CDN_SEGMENT, + confidence, RiskLevel.LOW, evidenceLines, RemediationKind.HOST_BLACKLIST); + } + + /** 区间内不属于 playlist 域名的切片 host，去重保序。 */ + private static Set foreignHosts(AdIntervalEvidence evidence) { + Set hosts = new LinkedHashSet<>(); + String playlistHost = evidence.playlistHost(); + for (SegmentFact fact : evidence.inside()) { + if (fact.host().isEmpty()) continue; + if (!playlistHost.isEmpty() && fact.hostEndsWith(playlistHost)) continue; + hosts.add(fact.host().toLowerCase(Locale.US)); + } + return hosts; + } + + private static List matching(Set hosts, List patterns) { + List matched = new ArrayList<>(); + for (String host : hosts) { + if (endsWithAny(host, patterns)) matched.add(host); + } + return matched; + } + + /** + * host 是否命中任一片段。黑名单条目多为域名片段（如 {@code doubleclick.net}）， + * 沿用现有拦截语义：后缀匹配或包含匹配。 + */ + private static boolean endsWithAny(String host, List patterns) { + String lower = host.toLowerCase(Locale.US); + for (String pattern : patterns) { + if (pattern == null || pattern.isBlank()) continue; + String target = pattern.toLowerCase(Locale.US).trim(); + if (lower.equals(target) || lower.endsWith("." + target) || lower.contains(target)) return true; + } + return false; + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java new file mode 100644 index 0000000000..2bf296d0d3 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java @@ -0,0 +1,117 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.ArrayList; +import java.util.List; +import java.util.Locale; + +/** + * 诊断通道：回答「为什么现有机制没拦住」。 + * + *

本通道不产出新规则，但会给出成本最低的修复建议 —— 启用一条已存在 + * 但默认关闭的规则，永远优于新建规则。见设计文档第 7.3、8.1 节。 + */ +public final class ExistingRuleClassifier { + + public static final String CHANNEL_ID = "existing-rule"; + + /** 命中一条已存在但未启用的规则时的置信度。 */ + static final float CONFIDENCE_DISABLED_RULE = 0.75f; + + private ExistingRuleClassifier() { + } + + /** + * 一条内置/外部 HLS 规则的当前状态，对应 {@code HlsRuleConfig.Entry} 的子集。 + * + * @param key 复合状态键 + * @param id 规则 id + * @param name 规则名 + * @param enabled 当前是否启用 + * @param valid 编译是否通过 + * @param hostSuffixes 该规则针对的切片域名后缀 + */ + public record RuleState(String key, String id, String name, boolean enabled, + boolean valid, List hostSuffixes) { + + public RuleState { + key = key == null ? "" : key; + id = id == null ? "" : id; + name = name == null ? "" : name; + hostSuffixes = hostSuffixes == null ? List.of() : List.copyOf(hostSuffixes); + } + } + + /** + * 分类输入。 + * + * @param hlsRules 已知的 HLS 规则状态 + * @param legacyHeuristicActive 旧启发式引擎当次是否生效 + * @param protectingExcludes 可能误保护广告切片的正片保护正则 + */ + public record Input(List hlsRules, boolean legacyHeuristicActive, + List protectingExcludes) { + + public Input { + hlsRules = hlsRules == null ? List.of() : List.copyOf(hlsRules); + protectingExcludes = protectingExcludes == null ? List.of() : List.copyOf(protectingExcludes); + } + + public static Input empty() { + return new Input(List.of(), false, List.of()); + } + } + + /** 无诊断结论时返回 null。 */ + public static AdAttribution classify(AdIntervalEvidence evidence, Input input) { + if (evidence == null || evidence.inside().isEmpty()) return null; + Input safe = input == null ? Input.empty() : input; + + RuleState disabled = findDisabledMatch(evidence, safe.hlsRules()); + if (disabled != null) { + List lines = new ArrayList<>(); + lines.add(String.format(Locale.US, "已有规则「%s」覆盖该域名但当前未启用", + disabled.name().isEmpty() ? disabled.id() : disabled.name())); + lines.add("启用它即可拦掉这段广告，无需新建规则"); + return new AdAttribution(CHANNEL_ID, AdCategory.DISCONTINUITY_BLOCK, + CONFIDENCE_DISABLED_RULE, RiskLevel.LOW, lines, + RemediationKind.ENABLE_EXISTING_RULE); + } + + List lines = new ArrayList<>(); + if (evidence.handledByStructuredRule()) { + lines.add("该区间已被规则 " + String.join("、", evidence.alreadyRemovedByStructuredRuleIds()) + + " 处理，用户看到的广告来自其他来源"); + } + if (safe.legacyHeuristicActive()) { + lines.add("旧启发式引擎当次生效，其判定不依赖规则，可能与结构化结论冲突"); + } + if (!safe.protectingExcludes().isEmpty()) { + lines.add("存在 " + safe.protectingExcludes().size() + + " 条正片保护规则，可能阻止了这些切片被删除"); + } + RuleState invalid = findInvalid(safe.hlsRules()); + if (invalid != null) { + lines.add("规则「" + (invalid.name().isEmpty() ? invalid.id() : invalid.name()) + + "」编译失败，未参与本次净化"); + } + if (lines.isEmpty()) return null; + + return new AdAttribution(CHANNEL_ID, AdCategory.ALREADY_HANDLED, + 0.5f, RiskLevel.LOW, lines, RemediationKind.NONE); + } + + /** 找出一条覆盖本区间域名、有效但未启用的规则。 */ + private static RuleState findDisabledMatch(AdIntervalEvidence evidence, List rules) { + for (RuleState rule : rules) { + if (rule.enabled() || !rule.valid() || rule.hostSuffixes().isEmpty()) continue; + boolean covers = evidence.inside().stream() + .anyMatch(fact -> rule.hostSuffixes().stream().anyMatch(fact::hostEndsWith)); + if (covers) return rule; + } + return null; + } + + private static RuleState findInvalid(List rules) { + return rules.stream().filter(rule -> !rule.valid()).findFirst().orElse(null); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiterTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiterTest.java new file mode 100644 index 0000000000..b09c7c8de6 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiterTest.java @@ -0,0 +1,144 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.Arrays; +import java.util.List; + +public class AdAttributionArbiterTest { + + private static AdAttribution of(String channel, AdCategory category, float confidence, + RemediationKind kind) { + return new AdAttribution(channel, category, confidence, RiskLevel.MEDIUM, + List.of(channel + " 证据"), kind); + } + + @Test + public void prefersCheaperMechanismAtComparableConfidence() { + // 域名黑名单置信度略低，但成本与风险都更低 + AdAttribution hls = of("hls", AdCategory.DISCONTINUITY_BLOCK, 0.80f, + RemediationKind.HLS_STRUCTURED_RULE); + AdAttribution domain = of("domain", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.78f, + RemediationKind.HOST_BLACKLIST); + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(List.of(hls, domain)); + + assertEquals(RemediationKind.HOST_BLACKLIST, verdict.preferred().remediation()); + assertTrue(verdict.hasActionablePlan()); + assertEquals(1, verdict.alternatives().size()); + } + + @Test + public void enablingExistingRuleOutranksNewRuleEvenAtLowerConfidence() { + AdAttribution hls = of("hls", AdCategory.DISCONTINUITY_BLOCK, 0.85f, + RemediationKind.HLS_STRUCTURED_RULE); + AdAttribution existing = of("existing-rule", AdCategory.FIXED_DURATION_BLOCK, 0.75f, + RemediationKind.ENABLE_EXISTING_RULE); + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(List.of(hls, existing)); + + assertEquals(RemediationKind.ENABLE_EXISTING_RULE, verdict.preferred().remediation()); + } + + @Test + public void muchHigherConfidenceStillWinsOverCheaperMechanism() { + // 置信度占 0.7 权重，差距足够大时仍可翻盘 + AdAttribution hls = of("hls", AdCategory.DISCONTINUITY_BLOCK, 1.0f, + RemediationKind.HLS_STRUCTURED_RULE); + AdAttribution domain = of("domain", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.30f, + RemediationKind.HOST_BLACKLIST); + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(List.of(hls, domain)); + + assertEquals(RemediationKind.HLS_STRUCTURED_RULE, verdict.preferred().remediation()); + } + + @Test + public void mergesSameCategoryWithProbabilisticOr() { + AdAttribution first = of("hls", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.6f, + RemediationKind.HLS_STRUCTURED_RULE); + AdAttribution second = of("domain", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.5f, + RemediationKind.HOST_BLACKLIST); + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(List.of(first, second)); + + // 1 - (1-0.6)(1-0.5) = 0.8 + assertEquals(0.8f, verdict.preferred().confidence(), 0.0001f); + // 合并后取成本更低的机制 + assertEquals(RemediationKind.HOST_BLACKLIST, verdict.preferred().remediation()); + // 证据合并，通道名拼接 + assertEquals(2, verdict.preferred().evidence().size()); + assertTrue(verdict.preferred().channelId().contains("hls")); + assertTrue(verdict.preferred().channelId().contains("domain")); + assertTrue(verdict.alternatives().isEmpty()); + } + + @Test + public void alreadyHandledIsDemotedToDiagnostics() { + AdAttribution diagnosis = new AdAttribution("domain", AdCategory.ALREADY_HANDLED, 0.95f, + RiskLevel.LOW, List.of("已在黑名单中"), RemediationKind.NONE); + AdAttribution hls = of("hls", AdCategory.DISCONTINUITY_BLOCK, 0.4f, + RemediationKind.HLS_STRUCTURED_RULE); + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(List.of(diagnosis, hls)); + + // 置信度 0.95 远高于 0.4，但不可落地，只能进诊断 + assertEquals(RemediationKind.HLS_STRUCTURED_RULE, verdict.preferred().remediation()); + assertEquals(1, verdict.diagnostics().size()); + assertEquals(AdCategory.ALREADY_HANDLED, verdict.diagnostics().get(0).category()); + } + + @Test + public void diagnosticsOnlyFallsBackToSessionSkip() { + AdAttribution diagnosis = new AdAttribution("existing-rule", AdCategory.ALREADY_HANDLED, 0.5f, + RiskLevel.LOW, List.of("旧启发式引擎生效"), RemediationKind.NONE); + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(List.of(diagnosis)); + + assertEquals(RemediationKind.SESSION_SKIP_ONLY, verdict.preferred().remediation()); + assertEquals(AdCategory.IN_STREAM_BURNED_IN, verdict.preferred().category()); + assertFalse(verdict.hasActionablePlan()); + // 诊断证据被并入兜底结论，便于 UI 一次展示 + assertTrue(verdict.preferred().evidence().stream() + .anyMatch(line -> line.contains("旧启发式引擎生效"))); + } + + @Test + public void allChannelsAbstainYieldsEmptyVerdict() { + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate( + Arrays.asList(null, null)); + + assertNull(verdict.preferred()); + assertTrue(verdict.empty()); + assertFalse(verdict.hasActionablePlan()); + } + + @Test + public void handlesNullListAndSkipsNullEntries() { + assertTrue(AdAttributionArbiter.arbitrate(null).empty()); + + AdAttribution hls = of("hls", AdCategory.DISCONTINUITY_BLOCK, 0.6f, + RemediationKind.HLS_STRUCTURED_RULE); + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate( + Arrays.asList(null, hls, null)); + + assertEquals(RemediationKind.HLS_STRUCTURED_RULE, verdict.preferred().remediation()); + } + + @Test + public void mergedRiskTakesLowestOfGroup() { + AdAttribution medium = new AdAttribution("hls", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.5f, + RiskLevel.MEDIUM, List.of("a"), RemediationKind.HLS_STRUCTURED_RULE); + AdAttribution low = new AdAttribution("domain", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.5f, + RiskLevel.LOW, List.of("b"), RemediationKind.HOST_BLACKLIST); + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(List.of(medium, low)); + + assertEquals(RiskLevel.LOW, verdict.preferred().risk()); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java new file mode 100644 index 0000000000..f4000a7759 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java @@ -0,0 +1,109 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertNotNull; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.List; + +public class DomainReputationClassifierTest { + + private static final String PLAYLIST_HOST = "v.example.com"; + + private static AdIntervalEvidence evidenceWith(String... insideHosts) { + List inside = new java.util.ArrayList<>(); + for (int i = 0; i < insideHosts.length; i++) { + inside.add(new SegmentFact(i, insideHosts[i], "/seg/" + i + ".ts", 6.4, false)); + } + return new AdIntervalEvidence( + 10_000, 40_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + PLAYLIST_HOST, "/play/index.m3u8", true, + inside, List.of(new SegmentFact(9, PLAYLIST_HOST, "/seg/9.ts", 8.0, false)), + false, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + @Test + public void blacklistedHostYieldsAlreadyHandledDiagnosis() { + DomainReputationClassifier.Input input = new DomainReputationClassifier.Input( + List.of("doubleclick.net"), List.of(PLAYLIST_HOST), List.of(), ""); + + AdAttribution attribution = DomainReputationClassifier.classify( + evidenceWith("ads.doubleclick.net"), input); + + assertNotNull(attribution); + // 已在黑名单却仍被看到，说明拦截路径没覆盖播放器直连请求 + assertEquals(AdCategory.ALREADY_HANDLED, attribution.category()); + assertEquals(RemediationKind.NONE, attribution.remediation()); + // ALREADY_HANDLED 的得分强制归零 + assertEquals(0f, attribution.score(), 0.0001f); + } + + @Test + public void hostOutsideSiteBaselineIsThirdPartyCdn() { + DomainReputationClassifier.Input input = new DomainReputationClassifier.Input( + List.of(), List.of(PLAYLIST_HOST, "cdn.example.com"), List.of(), ""); + + AdAttribution attribution = DomainReputationClassifier.classify( + evidenceWith("ad-cdn.other.com"), input); + + assertNotNull(attribution); + assertEquals(AdCategory.THIRD_PARTY_CDN_SEGMENT, attribution.category()); + assertEquals(RemediationKind.HOST_BLACKLIST, attribution.remediation()); + assertEquals(RiskLevel.LOW, attribution.risk()); + assertEquals(DomainReputationClassifier.CONFIDENCE_UNKNOWN_HOST, + attribution.confidence(), 0.0001f); + } + + @Test + public void interfaceCandidateAddsConfidenceBonus() { + DomainReputationClassifier.Input input = new DomainReputationClassifier.Input( + List.of(), List.of(PLAYLIST_HOST), List.of("ad-cdn.other.com"), "饭太硬"); + + AdAttribution attribution = DomainReputationClassifier.classify( + evidenceWith("ad-cdn.other.com"), input); + + assertNotNull(attribution); + assertEquals(DomainReputationClassifier.CONFIDENCE_UNKNOWN_HOST + + DomainReputationClassifier.BONUS_INTERFACE_CANDIDATE, + attribution.confidence(), 0.0001f); + assertTrue(attribution.evidence().stream().anyMatch(line -> line.contains("饭太硬"))); + } + + @Test + public void abstainsWithoutBaselineToAvoidFirstPlayMisjudgement() { + // 首次播放该站，没有任何基线数据：不能断言域名异常 + DomainReputationClassifier.Input input = DomainReputationClassifier.Input.empty(); + + assertNull(DomainReputationClassifier.classify(evidenceWith("ad-cdn.other.com"), input)); + } + + @Test + public void abstainsWhenAllSegmentsBelongToPlaylistHost() { + DomainReputationClassifier.Input input = new DomainReputationClassifier.Input( + List.of(), List.of(PLAYLIST_HOST), List.of(), ""); + + assertNull(DomainReputationClassifier.classify(evidenceWith(PLAYLIST_HOST), input)); + // 子域名同样视为本站 + assertNull(DomainReputationClassifier.classify(evidenceWith("cdn." + PLAYLIST_HOST), input)); + } + + @Test + public void abstainsWhenForeignHostIsInBaseline() { + // 该站正常使用独立 CDN 域名，已在基线内 + DomainReputationClassifier.Input input = new DomainReputationClassifier.Input( + List.of(), List.of(PLAYLIST_HOST, "fast-cdn.other.com"), List.of(), ""); + + assertNull(DomainReputationClassifier.classify(evidenceWith("fast-cdn.other.com"), input)); + } + + @Test + public void handlesNullInputAndEmptyInterval() { + assertNull(DomainReputationClassifier.classify(null, DomainReputationClassifier.Input.empty())); + assertNull(DomainReputationClassifier.classify(evidenceWith("ad.other.com"), null)); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifierTest.java new file mode 100644 index 0000000000..bee31c8d19 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifierTest.java @@ -0,0 +1,112 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertNotNull; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.List; + +public class ExistingRuleClassifierTest { + + private static final String PLAYLIST_HOST = "v.example.com"; + + private static AdIntervalEvidence evidence(String insideHost, List removedRuleIds) { + return new AdIntervalEvidence( + 10_000, 40_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + PLAYLIST_HOST, "/play/index.m3u8", true, + List.of(new SegmentFact(3, insideHost, "/seg/3.ts", 6.4, true)), + List.of(new SegmentFact(0, PLAYLIST_HOST, "/seg/0.ts", 8.0, false)), + false, false, removedRuleIds, false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + @Test + public void suggestsEnablingDisabledRuleInsteadOfCreatingNew() { + ExistingRuleClassifier.RuleState disabled = new ExistingRuleClassifier.RuleState( + "builtin:baofeng", "baofeng-preroll", "暴风片头", false, true, + List.of("ad-cdn.other.com")); + + AdAttribution attribution = ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of()), + new ExistingRuleClassifier.Input(List.of(disabled), false, List.of())); + + assertNotNull(attribution); + assertEquals(RemediationKind.ENABLE_EXISTING_RULE, attribution.remediation()); + assertEquals(ExistingRuleClassifier.CONFIDENCE_DISABLED_RULE, + attribution.confidence(), 0.0001f); + assertTrue(attribution.actionable()); + assertTrue(attribution.evidence().stream().anyMatch(line -> line.contains("暴风片头"))); + } + + @Test + public void ignoresAlreadyEnabledRule() { + ExistingRuleClassifier.RuleState enabled = new ExistingRuleClassifier.RuleState( + "builtin:baofeng", "baofeng-preroll", "暴风片头", true, true, + List.of("ad-cdn.other.com")); + + assertNull(ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of()), + new ExistingRuleClassifier.Input(List.of(enabled), false, List.of()))); + } + + @Test + public void ignoresDisabledRuleThatDoesNotCoverInterval() { + ExistingRuleClassifier.RuleState unrelated = new ExistingRuleClassifier.RuleState( + "builtin:other", "other-rule", "其他规则", false, true, + List.of("unrelated-cdn.com")); + + assertNull(ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of()), + new ExistingRuleClassifier.Input(List.of(unrelated), false, List.of()))); + } + + @Test + public void reportsInvalidRuleAsDiagnosis() { + ExistingRuleClassifier.RuleState invalid = new ExistingRuleClassifier.RuleState( + "builtin:broken", "broken-rule", "坏规则", false, false, List.of("ad-cdn.other.com")); + + AdAttribution attribution = ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of()), + new ExistingRuleClassifier.Input(List.of(invalid), false, List.of())); + + assertNotNull(attribution); + // 编译失败的规则不能建议启用，只能报诊断 + assertEquals(RemediationKind.NONE, attribution.remediation()); + assertEquals(AdCategory.ALREADY_HANDLED, attribution.category()); + assertTrue(attribution.evidence().stream().anyMatch(line -> line.contains("编译失败"))); + } + + @Test + public void reportsLegacyHeuristicAndProtectingExcludes() { + AdAttribution attribution = ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of()), + new ExistingRuleClassifier.Input(List.of(), true, List.of(".*/main/.*"))); + + assertNotNull(attribution); + assertEquals(RemediationKind.NONE, attribution.remediation()); + assertTrue(attribution.evidence().stream().anyMatch(line -> line.contains("旧启发式引擎"))); + assertTrue(attribution.evidence().stream().anyMatch(line -> line.contains("正片保护"))); + } + + @Test + public void reportsIntervalAlreadyRemovedByStructuredRule() { + AdAttribution attribution = ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of("quantum-block")), + ExistingRuleClassifier.Input.empty()); + + assertNotNull(attribution); + assertTrue(attribution.evidence().stream().anyMatch(line -> line.contains("quantum-block"))); + } + + @Test + public void abstainsWhenNothingToReport() { + assertNull(ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of()), + ExistingRuleClassifier.Input.empty())); + assertNull(ExistingRuleClassifier.classify(null, ExistingRuleClassifier.Input.empty())); + } +} From 2f1b919325f879be8f262110ff7a66ad9afb0185 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sat, 29 Aug 2026 23:03:17 +0800 Subject: [PATCH 04/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9AAdSkipCoordina?= =?UTF-8?q?tor=20=E6=94=AF=E6=8C=81=E7=94=A8=E6=88=B7=E6=A1=86=E9=80=89?= =?UTF-8?q?=E5=8C=BA=E9=97=B4=E7=9A=84=E5=8D=B3=E6=97=B6=E8=B7=B3=E8=BF=87?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第四个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 9 节。 onUserInterval() 与音频候选路径的两个关键区别： - 区间是媒体时间，直接用作 seek 目标，不经 clock().mapCaptureToMediaMs() 二次映射（音频候选的 endTimeMs 是采集时间，必须映射）。 - 授权来自用户框选动作本身，不经 PROMPT_PENDING 确认，直接进 SEEKING。 安全校验与撤销窗口保持一致：session/generation 匹配、可 seek、非直播、 时长已知；终点超出时长则钳制；终点已播过（短按场景广告已放完）则记录 SEEK_REJECTED 并返回 false，由调用方提示「已记录，未跳过」；撤销窗口内 同一区间重复提交不再 seek。 seek 失败时恢复调用前状态而非 clearState()：音频候选的 suppressedKey 与 PROMPT_PENDING 不属于本次区间反馈，连带清掉会让同一广告再次弹提示。 音频路径不受此影响（onAutoCandidate 进 seek 前已主动清 suppressedKey）， 这条只在新增的用户区间路径上成立，已由测试锁定。 Verification: testMobileArm64_v8aDebugUnitTest --tests com.fongmi.android.tv.ad.* : 37 suites, 332 tests, 0 failures (AdSkipCoordinatorUserIntervalTest 15 new, AdSkipCoordinatorTest 9 unchanged) Task-Guard: AF3-user-interval-skip --- .../tv/ad/audio/AdSkipCoordinator.java | 60 ++++ .../AdSkipCoordinatorUserIntervalTest.java | 339 ++++++++++++++++++ 2 files changed, 399 insertions(+) create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/audio/AdSkipCoordinatorUserIntervalTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/audio/AdSkipCoordinator.java b/app/src/main/java/com/fongmi/android/tv/ad/audio/AdSkipCoordinator.java index 6087965b8f..23836b5441 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/audio/AdSkipCoordinator.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/audio/AdSkipCoordinator.java @@ -147,6 +147,66 @@ public final class AdSkipCoordinator implements AutoCloseable { return applyAutomaticSeek(legacy, target); } + /** + * 用户显式框选的广告区间。与 {@link #onCandidate} 的区别是授权来源： + * 框选动作本身即为对本次跳过的授权，因此不经 PROMPT_PENDING 直接执行， + * 但全部安全校验照旧，并保留撤销窗口。 + * + *

{@code startMs} / {@code endMs} 已经是媒体时间，不做采集时钟映射。 + * + * @return 是否真的执行了 seek。终点已过或校验失败时返回 false， + * 调用方应把它当作「已记录，未跳过」而不是失败。 + */ + public synchronized boolean onUserInterval(long sessionId, long generation, + long startMs, long endMs, String feedbackId) { + Objects.requireNonNull(feedbackId, "feedbackId"); + if (closed) return false; + if (endMs <= startMs || startMs < 0L) return false; + if (state == State.SEEKING || state == State.UNDO_WINDOW) return false; + + CandidateKey key = new CandidateKey(sessionId, generation, feedbackId, startMs, endMs); + // 连按去重：同一区间不重复 seek + if (key.equals(suppressedKey) || key.equals(activeKey)) return false; + + PlaybackSnapshot snapshot = playback.snapshot(sessionId, generation); + if (!isCurrentSeekable(snapshot, sessionId, generation)) { + diagnostics.record(AdAudioDiagnostics.Code.SEEK_REJECTED); + return false; + } + long target = clamp(endMs, 0L, snapshot.durationMs()); + // 终点已经播过去了，没有可跳的内容 + if (target <= snapshot.positionMs()) { + diagnostics.record(AdAudioDiagnostics.Code.SEEK_REJECTED); + return false; + } + + long originalPositionMs = snapshot.positionMs(); + // 失败时要还原到调用前的状态：音频候选的 suppressedKey / PROMPT_PENDING + // 不属于本次区间反馈，不能被连带清掉。 + State previousState = state; + state = State.SEEKING; + SeekResult seek = playback.seekTo(sessionId, generation, target); + if (seek == null || !seek.applied() || seek.sessionId() != sessionId) { + diagnostics.record(AdAudioDiagnostics.Code.SEEK_REJECTED); + state = previousState; + return false; + } + + long deadline = saturatedAdd(elapsedRealtime.getAsLong(), undoWindowMs); + undoContext = new UndoContext(sessionId, seek.generation(), feedbackId, + originalPositionMs, target, deadline); + activeCandidate = null; + activeKey = null; + state = State.UNDO_WINDOW; + long undoToken = ++actionToken; + AdAudioDiagnostics.log("user interval skip id=%s targetMs=%d fromMs=%d", + feedbackId, target, originalPositionMs); + ui.showUndo(new UndoPrompt(undoContext.sessionId, undoContext.generation, + undoContext.ruleId, undoContext.originalPositionMs, + undoContext.targetPositionMs, undoWindowMs), new TokenActions(undoToken)); + return true; + } + public synchronized void reset(long sessionId) { if (closed) return; actionToken++; diff --git a/app/src/test/java/com/fongmi/android/tv/ad/audio/AdSkipCoordinatorUserIntervalTest.java b/app/src/test/java/com/fongmi/android/tv/ad/audio/AdSkipCoordinatorUserIntervalTest.java new file mode 100644 index 0000000000..742b4d6851 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/audio/AdSkipCoordinatorUserIntervalTest.java @@ -0,0 +1,339 @@ +package com.fongmi.android.tv.ad.audio; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import com.fongmi.android.tv.player.audio.PlaybackMediaClock; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; +import java.util.function.LongSupplier; + +/** + * {@link AdSkipCoordinator#onUserInterval} 的行为锁定。 + * + *

与音频候选路径的关键区别：用户区间是媒体时间，不经采集时钟映射； + * 授权来自用户框选，不经 PROMPT_PENDING 确认。 + */ +public class AdSkipCoordinatorUserIntervalTest { + + @Test + public void skipsImmediatelyWithoutPromptAndKeepsUndo() { + ManualTime time = new ManualTime(); + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 135_000L, 1_200_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L, time); + + boolean applied = coordinator.onUserInterval(7L, 2L, 135_000L, 165_000L, "fb-1"); + + assertTrue(applied); + // 不经确认对话框 + assertEquals(0, ui.candidateShows); + assertEquals(1, ui.undoShows); + assertEquals(List.of(165_000L), playback.seekTargets); + assertEquals(AdSkipCoordinator.State.UNDO_WINDOW, coordinator.state()); + } + + @Test + public void undoReturnsToPositionBeforeSkip() { + ManualTime time = new ManualTime(); + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 135_000L, 1_200_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L, time); + + coordinator.onUserInterval(7L, 2L, 135_000L, 165_000L, "fb-1"); + ui.actions.undo(); + + assertEquals(List.of(165_000L, 135_000L), playback.seekTargets); + assertEquals(AdSkipCoordinator.State.IDLE, coordinator.state()); + } + + @Test + public void usesMediaTimeDirectlyWithoutCaptureClockMapping() { + // 采集时钟严重滞后于播放位置。若误走 mapCaptureToMediaMs 会得到别的目标值。 + ManualTime time = new ManualTime(); + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 1L, 1L, 100_000L, 600_000L, true, false, + new PlaybackMediaClock.Snapshot(1L, 0L, 20_000L, 15_000L, true, true))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L, time); + + assertTrue(coordinator.onUserInterval(1L, 1L, 100_000L, 130_000L, "fb-1")); + + // 终点原样使用 + assertEquals(List.of(130_000L), playback.seekTargets); + } + + @Test + public void staleGenerationIsRejected() { + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 3L, 10_000L, 600_000L, true, false, freshClock(3L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + assertFalse(coordinator.onUserInterval(7L, 2L, 10_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + assertEquals(AdSkipCoordinator.State.IDLE, coordinator.state()); + } + + @Test + public void mismatchedSessionIsRejected() { + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 10_000L, 600_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + assertFalse(coordinator.onUserInterval(8L, 2L, 10_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + } + + @Test + public void livePlaybackIsRejected() { + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 10_000L, 600_000L, true, true, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + assertFalse(coordinator.onUserInterval(7L, 2L, 10_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + } + + @Test + public void unseekablePlaybackIsRejected() { + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 10_000L, 600_000L, false, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + assertFalse(coordinator.onUserInterval(7L, 2L, 10_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + } + + @Test + public void intervalAlreadyPlayedPastIsRecordedButNotSkipped() { + // 短按场景：广告已经放完，当前位置在区间终点之后 + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 200_000L, 600_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + assertFalse(coordinator.onUserInterval(7L, 2L, 135_000L, 165_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + assertEquals(AdSkipCoordinator.State.IDLE, coordinator.state()); + } + + @Test + public void endBeyondDurationIsClamped() { + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 580_000L, 600_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + assertTrue(coordinator.onUserInterval(7L, 2L, 580_000L, 900_000L, "fb-1")); + + assertEquals(List.of(600_000L), playback.seekTargets); + } + + @Test + public void duplicateSubmissionDoesNotSeekTwice() { + ManualTime time = new ManualTime(); + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 135_000L, 1_200_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L, time); + + assertTrue(coordinator.onUserInterval(7L, 2L, 135_000L, 165_000L, "fb-1")); + // 撤销窗口内连按第二次 + assertFalse(coordinator.onUserInterval(7L, 2L, 135_000L, 165_000L, "fb-1")); + + assertEquals(List.of(165_000L), playback.seekTargets); + } + + @Test + public void invalidIntervalIsRejected() { + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 10_000L, 600_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + // 终点不晚于起点 + assertFalse(coordinator.onUserInterval(7L, 2L, 40_000L, 40_000L, "fb-1")); + assertFalse(coordinator.onUserInterval(7L, 2L, 50_000L, 40_000L, "fb-1")); + // 负起点 + assertFalse(coordinator.onUserInterval(7L, 2L, -1_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + } + + @Test + public void closedCoordinatorRejectsInterval() { + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 10_000L, 600_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + coordinator.close(); + + assertFalse(coordinator.onUserInterval(7L, 2L, 10_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + } + + @Test + public void undoAfterDeadlineDoesNotSeekBack() { + ManualTime time = new ManualTime(); + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 7L, 2L, 135_000L, 1_200_000L, true, false, freshClock(2L))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L, time); + + coordinator.onUserInterval(7L, 2L, 135_000L, 165_000L, "fb-1"); + time.now = 5_001L; + ui.actions.undo(); + + assertEquals(List.of(165_000L), playback.seekTargets); + assertEquals(AdSkipCoordinator.State.IDLE, coordinator.state()); + } + + @Test + public void audioCandidatePromptIsNotDisturbedByRejectedInterval() { + // 区间被拒不应清掉正在展示的音频候选提示 + FakePlaybackPort playback = new FakePlaybackPort(snapshot( + 1L, 1L, 5_000L, 100_000L, true, false, + new PlaybackMediaClock.Snapshot(1L, 0L, 30_000L, 5_000L, true, true))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + coordinator.onCandidate(new AdAudioConsumer.Candidate(1L, 1L, + new AudioFingerprintMatcher.MatchEvent( + AudioFingerprintMatcher.Type.START_MATCHED, "ad-1", 0L, 10_000L, 0.9f, 6))); + assertEquals(AdSkipCoordinator.State.PROMPT_PENDING, coordinator.state()); + + // 用无效区间触发拒绝 + assertFalse(coordinator.onUserInterval(1L, 1L, 40_000L, 40_000L, "fb-1")); + + assertEquals(AdSkipCoordinator.State.PROMPT_PENDING, coordinator.state()); + assertEquals("ad-1", ui.lastPrompt.ruleId()); + } + + @Test + public void failedSeekDoesNotClearSuppressedAudioCandidate() { + // 音频候选被 ignore 后，其 suppressedKey 必须存活； + // 用户区间 seek 失败不应把它一起清掉，否则同一广告会再次弹提示。 + RejectingSeekPort playback = new RejectingSeekPort(snapshot( + 1L, 1L, 5_000L, 100_000L, true, false, + new PlaybackMediaClock.Snapshot(1L, 0L, 30_000L, 5_000L, true, true))); + FakeUiPort ui = new FakeUiPort(); + AdSkipCoordinator coordinator = new AdSkipCoordinator(playback, ui, 5_000L); + + AdAudioConsumer.Candidate audio = new AdAudioConsumer.Candidate(1L, 1L, + new AudioFingerprintMatcher.MatchEvent( + AudioFingerprintMatcher.Type.START_MATCHED, "ad-1", 0L, 10_000L, 0.9f, 6)); + coordinator.onCandidate(audio); + ui.actions.ignore(); + assertEquals(AdSkipCoordinator.State.SUPPRESSED, coordinator.state()); + + // 用户区间通过校验但 seekTo 被拒 + assertFalse(coordinator.onUserInterval(1L, 1L, 5_000L, 40_000L, "fb-1")); + + // 同一音频候选不得再次弹出 + int showsBefore = ui.candidateShows; + coordinator.onCandidate(audio); + assertEquals(showsBefore, ui.candidateShows); + } + + private static AdSkipCoordinator.PlaybackSnapshot snapshot( + long session, long generation, long position, long duration, + boolean seekable, boolean live, PlaybackMediaClock.Snapshot clock) { + return new AdSkipCoordinator.PlaybackSnapshot( + session, generation, position, duration, seekable, live, clock); + } + + private static PlaybackMediaClock.Snapshot freshClock(long generation) { + return new PlaybackMediaClock.Snapshot(generation, 0L, 0L, 0L, true, true); + } + + private static final class ManualTime implements LongSupplier { + long now; + + @Override + public long getAsLong() { + return now; + } + } + + private static final class FakePlaybackPort implements AdSkipCoordinator.PlaybackPort { + private AdSkipCoordinator.PlaybackSnapshot snapshot; + private final List seekTargets = new ArrayList<>(); + + FakePlaybackPort(AdSkipCoordinator.PlaybackSnapshot snapshot) { + this.snapshot = snapshot; + } + + @Override + public AdSkipCoordinator.PlaybackSnapshot snapshot(long sessionId, long generation) { + return snapshot; + } + + @Override + public AdSkipCoordinator.SeekResult seekTo(long sessionId, long generation, long positionMs) { + if (snapshot.sessionId() != sessionId || snapshot.generation() != generation) { + return AdSkipCoordinator.SeekResult.rejected(sessionId, generation); + } + seekTargets.add(positionMs); + long nextGeneration = generation == Long.MAX_VALUE ? 0L : generation + 1L; + PlaybackMediaClock.Snapshot nextClock = new PlaybackMediaClock.Snapshot( + nextGeneration, snapshot.clock().mediaAnchorMs(), snapshot.clock().capturedUntilMs(), + snapshot.clock().presentedCaptureMs(), snapshot.clock().playing(), snapshot.clock().fresh()); + snapshot = new AdSkipCoordinator.PlaybackSnapshot( + snapshot.sessionId(), nextGeneration, positionMs, snapshot.durationMs(), + snapshot.seekable(), snapshot.live(), nextClock); + return new AdSkipCoordinator.SeekResult(true, sessionId, nextGeneration); + } + } + + /** 快照校验通过，但 seekTo 一律拒绝。 */ + private static final class RejectingSeekPort implements AdSkipCoordinator.PlaybackPort { + private final AdSkipCoordinator.PlaybackSnapshot snapshot; + + RejectingSeekPort(AdSkipCoordinator.PlaybackSnapshot snapshot) { + this.snapshot = snapshot; + } + + @Override + public AdSkipCoordinator.PlaybackSnapshot snapshot(long sessionId, long generation) { + return snapshot; + } + + @Override + public AdSkipCoordinator.SeekResult seekTo(long sessionId, long generation, long positionMs) { + return AdSkipCoordinator.SeekResult.rejected(sessionId, generation); + } + } + + private static final class FakeUiPort implements AdSkipCoordinator.UiPort { + private int candidateShows; + private int undoShows; + private AdSkipCoordinator.Prompt lastPrompt; + private AdSkipCoordinator.Actions actions; + + @Override + public void showCandidate(AdSkipCoordinator.Prompt prompt, AdSkipCoordinator.Actions actions) { + candidateShows++; + lastPrompt = prompt; + this.actions = actions; + } + + @Override + public void showUndo(AdSkipCoordinator.UndoPrompt prompt, AdSkipCoordinator.Actions actions) { + undoShows++; + this.actions = actions; + } + + @Override + public void dismiss(long sessionId) { + } + } +} From 7c843fd05fba3a28f6f741cee04028b48e895671 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sat, 29 Aug 2026 23:13:17 +0800 Subject: [PATCH 05/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E7=AB=99?= =?UTF-8?q?=E7=82=B9=E5=9F=9F=E5=90=8D=E5=9F=BA=E7=BA=BF=E3=80=81=E5=8F=8D?= =?UTF-8?q?=E9=A6=88=E4=BC=9A=E8=AF=9D=E7=8A=B6=E6=80=81=E4=B8=8E=E8=AF=8A?= =?UTF-8?q?=E6=96=AD=E8=AE=A1=E6=95=B0?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第五个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 8.3、11 节。 - SitePlaylistHostBaseline：每站点最近 8 个域名、全局 200 站点的双层 LRU， 供 DomainReputationClassifier 判断「非当前适配的域名」。只存 host 字符串， 不存 URL 全文、query 或 token。持久化经 Storage 接口注入，紧凑文本格式 而非 JSON，避免依赖 Gson 以便纯 JVM 单测；siteKey 中的分隔符做转义。 - AdFeedbackSession：驱动对话框两屏的状态，withVerdict() 从第一屏推进到第二屏。 - AdFeedbackDiagnostics：12 个固定枚举计数，沿用 AdAudioDiagnostics 的 「不记录 URL、规则正文、Cookie、Authorization」约定。 修了一处 record() 的缺陷：全部输入为空白时 computeIfAbsent 会留下空条目 白占站点配额，现改为移除。尚未接线到播放器，无行为变化。 Verification: testMobileArm64_v8aDebugUnitTest --tests com.fongmi.android.tv.ad.feedback.* : 59 tests (10 new baseline), 0 failures, 0 errors Task-Guard: AF4a-feedback-controller --- .../tv/ad/feedback/AdFeedbackDiagnostics.java | 67 ++++++++ .../tv/ad/feedback/AdFeedbackSession.java | 55 +++++++ .../ad/feedback/SitePlaylistHostBaseline.java | 150 ++++++++++++++++++ .../SitePlaylistHostBaselineTest.java | 140 ++++++++++++++++ 4 files changed, 412 insertions(+) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDiagnostics.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackSession.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaselineTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDiagnostics.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDiagnostics.java new file mode 100644 index 0000000000..b468ae3669 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDiagnostics.java @@ -0,0 +1,67 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.github.catvod.crawler.SpiderDebug; + +import java.util.Collections; +import java.util.EnumMap; +import java.util.Map; + +/** + * 区间反馈的本地诊断计数。只暴露固定枚举计数，不记录媒体 URL、规则正文、 + * Cookie、Authorization 或用户标识 —— 与 {@code AdAudioDiagnostics} 同一约定。 + */ +public final class AdFeedbackDiagnostics { + + static final String LOG_TAG = "ad-feedback"; + + public enum Code { + INTERVAL_SUBMITTED, + START_INFERRED_FROM_DISCONTINUITY, + START_INFERRED_FROM_FALLBACK, + IMMEDIATE_SKIP_APPLIED, + IMMEDIATE_SKIP_REJECTED, + EVIDENCE_COLLECT_FAILED, + CHANNEL_ABSTAINED, + ARBITER_NO_PLAN, + PLAN_ACCEPTED, + PLAN_DISCARDED, + DUPLICATE_SUBMISSION, + ALREADY_HANDLED_DETECTED + } + + private final EnumMap counts = new EnumMap<>(Code.class); + private Code lastCode; + + public void record(Code code) { + if (code == null) return; + synchronized (this) { + counts.put(code, counts.getOrDefault(code, 0L) + 1L); + lastCode = code; + } + } + + /** 把起点来源折算成对应的诊断码。 */ + public void recordStartOrigin(StartOrigin origin) { + if (origin == StartOrigin.DISCONTINUITY) { + record(Code.START_INFERRED_FROM_DISCONTINUITY); + } else if (origin == StartOrigin.FALLBACK_WINDOW) { + record(Code.START_INFERRED_FROM_FALLBACK); + } + } + + public synchronized long count(Code code) { + return counts.getOrDefault(code, 0L); + } + + public synchronized Code lastCode() { + return lastCode; + } + + public synchronized Map snapshot() { + return Collections.unmodifiableMap(new EnumMap<>(counts)); + } + + static void log(String message, Object... args) { + SpiderDebug.log(LOG_TAG + " " + String.format(message, args)); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackSession.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackSession.java new file mode 100644 index 0000000000..b0e9f65254 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackSession.java @@ -0,0 +1,55 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.List; + +/** + * 一次区间反馈的完整状态，驱动对话框两屏展示。 + * + * @param feedbackId 去重用的稳定标识 + * @param startMs 区间起点 + * @param endMs 区间终点 + * @param startOrigin 起点来源，第一屏需明示 + * @param skipApplied 本次播放是否已跳过 + * @param verdict 归因仲裁结果，第二屏内容；尚未完成时为 null + */ +public record AdFeedbackSession(String feedbackId, long startMs, long endMs, + StartOrigin startOrigin, boolean skipApplied, + AdAttributionArbiter.Verdict verdict) { + + public AdFeedbackSession { + if (feedbackId == null || feedbackId.isBlank()) { + throw new IllegalArgumentException("feedbackId is required"); + } + startOrigin = startOrigin == null ? StartOrigin.FALLBACK_WINDOW : startOrigin; + } + + /** 第一屏：区间已确定，归因未开始。 */ + public static AdFeedbackSession pending(String feedbackId, long startMs, long endMs, + StartOrigin startOrigin, boolean skipApplied) { + return new AdFeedbackSession(feedbackId, startMs, endMs, startOrigin, skipApplied, null); + } + + /** 第二屏：附加归因结果。 */ + public AdFeedbackSession withVerdict(AdAttributionArbiter.Verdict verdict) { + return new AdFeedbackSession(feedbackId, startMs, endMs, startOrigin, skipApplied, verdict); + } + + public long durationMs() { + return endMs - startMs; + } + + public boolean analysisComplete() { + return verdict != null; + } + + /** 有可落地的规则方案。 */ + public boolean hasActionablePlan() { + return verdict != null && verdict.hasActionablePlan(); + } + + /** 首选方案的证据，无结论时为空。 */ + public List preferredEvidence() { + if (verdict == null || verdict.preferred() == null) return List.of(); + return verdict.preferred().evidence(); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java new file mode 100644 index 0000000000..21fd6ccabf --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java @@ -0,0 +1,150 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.ArrayList; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Locale; +import java.util.Map; + +/** + * 每站点最近成功播放的域名基线，供 {@link DomainReputationClassifier} 判断 + * 「非当前适配的域名」。见设计文档第 8.3 节。 + * + *

只存 host 字符串，不存 URL 全文、query 或 token。单站点上限 8 个域名， + * 全局上限 200 个站点，均按最久未用淘汰。 + * + *

本类只负责内存中的 LRU 结构与序列化，持久化由调用方注入的 + * {@link Storage} 完成，因此可在纯 JVM 下单测。 + */ +public final class SitePlaylistHostBaseline { + + /** 单站点保留的域名数上限。 */ + public static final int MAX_HOSTS_PER_SITE = 8; + /** 全局保留的站点数上限。 */ + public static final int MAX_SITES = 200; + + /** 持久化后端，实现通常包装 SharedPreferences。 */ + public interface Storage { + String read(); + + void write(String value); + } + + /** 内存实现，测试与无持久化场景使用。 */ + public static final class MemoryStorage implements Storage { + private String value = ""; + + @Override + public String read() { + return value; + } + + @Override + public void write(String value) { + this.value = value == null ? "" : value; + } + } + + private final Storage storage; + // accessOrder=true：get 也算一次使用，实现真正的 LRU + private final LinkedHashMap> sites = + new LinkedHashMap<>(16, 0.75f, true); + + public SitePlaylistHostBaseline(Storage storage) { + this.storage = storage == null ? new MemoryStorage() : storage; + load(); + } + + /** 记录一次成功播放涉及的域名。 */ + public synchronized void record(String siteKey, String... hosts) { + if (siteKey == null || siteKey.isBlank() || hosts == null) return; + List known = sites.computeIfAbsent(siteKey, key -> new ArrayList<>()); + boolean changed = false; + for (String host : hosts) { + if (host == null || host.isBlank()) continue; + String normalized = host.toLowerCase(Locale.US).trim(); + // 已存在则移到末尾，表示最近使用 + known.remove(normalized); + known.add(normalized); + changed = true; + } + if (!changed) { + // 全部输入为空白：不能留下空条目占用站点配额 + if (known.isEmpty()) sites.remove(siteKey); + return; + } + while (known.size() > MAX_HOSTS_PER_SITE) known.remove(0); + evictSites(); + persist(); + } + + /** 某站点已知的域名，最近使用的在后。 */ + public synchronized List hosts(String siteKey) { + if (siteKey == null) return List.of(); + List known = sites.get(siteKey); + return known == null ? List.of() : List.copyOf(known); + } + + public synchronized int siteCount() { + return sites.size(); + } + + public synchronized void clear() { + sites.clear(); + persist(); + } + + private void evictSites() { + // LinkedHashMap 的迭代顺序即访问顺序，最久未用在前 + while (sites.size() > MAX_SITES) { + String oldest = sites.keySet().iterator().next(); + sites.remove(oldest); + } + } + + /** + * 序列化为紧凑文本：{@code siteKey|host1,host2;siteKey2|host3}。 + * 不用 JSON 是为了避免依赖 Gson，本类需在纯 JVM 下可测。 + */ + private void persist() { + StringBuilder builder = new StringBuilder(); + for (Map.Entry> entry : sites.entrySet()) { + if (entry.getValue().isEmpty()) continue; + if (builder.length() > 0) builder.append(';'); + builder.append(escape(entry.getKey())).append('|'); + builder.append(String.join(",", entry.getValue())); + } + storage.write(builder.toString()); + } + + private void load() { + String raw = storage.read(); + if (raw == null || raw.isBlank()) return; + for (String chunk : raw.split(";")) { + int separator = chunk.indexOf('|'); + if (separator <= 0) continue; + String siteKey = unescape(chunk.substring(0, separator)); + String hostList = chunk.substring(separator + 1); + if (siteKey.isBlank() || hostList.isBlank()) continue; + List hosts = new ArrayList<>(); + for (String host : hostList.split(",")) { + if (!host.isBlank()) hosts.add(host.trim()); + } + if (hosts.isEmpty()) continue; + while (hosts.size() > MAX_HOSTS_PER_SITE) hosts.remove(0); + sites.put(siteKey, hosts); + } + evictSites(); + } + + /** siteKey 可能含分隔符，转义避免破坏格式。 */ + private static String escape(String value) { + return value.replace("%", "%25").replace(";", "%3B") + .replace("|", "%7C").replace(",", "%2C"); + } + + private static String unescape(String value) { + return value.replace("%2C", ",").replace("%7C", "|") + .replace("%3B", ";").replace("%25", "%"); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaselineTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaselineTest.java new file mode 100644 index 0000000000..e1b2bc828b --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaselineTest.java @@ -0,0 +1,140 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.List; + +public class SitePlaylistHostBaselineTest { + + @Test + public void recordsAndReadsBackHosts() { + SitePlaylistHostBaseline baseline = + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage()); + + baseline.record("siteA", "v.example.com", "cdn.example.com"); + + assertEquals(List.of("v.example.com", "cdn.example.com"), baseline.hosts("siteA")); + } + + @Test + public void normalisesHostCaseAndWhitespace() { + SitePlaylistHostBaseline baseline = + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage()); + + baseline.record("siteA", " V.Example.COM "); + + assertEquals(List.of("v.example.com"), baseline.hosts("siteA")); + } + + @Test + public void repeatedHostMovesToMostRecentWithoutDuplicating() { + SitePlaylistHostBaseline baseline = + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage()); + + baseline.record("siteA", "a.com", "b.com"); + baseline.record("siteA", "a.com"); + + assertEquals(List.of("b.com", "a.com"), baseline.hosts("siteA")); + } + + @Test + public void evictsOldestHostBeyondPerSiteLimit() { + SitePlaylistHostBaseline baseline = + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage()); + + for (int i = 0; i < SitePlaylistHostBaseline.MAX_HOSTS_PER_SITE + 3; i++) { + baseline.record("siteA", "host" + i + ".com"); + } + + List hosts = baseline.hosts("siteA"); + assertEquals(SitePlaylistHostBaseline.MAX_HOSTS_PER_SITE, hosts.size()); + // 最早的三个被淘汰 + assertTrue(hosts.contains("host3.com")); + assertTrue(hosts.stream().noneMatch(host -> host.equals("host0.com"))); + } + + @Test + public void evictsLeastRecentlyUsedSiteBeyondGlobalLimit() { + SitePlaylistHostBaseline baseline = + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage()); + + for (int i = 0; i < SitePlaylistHostBaseline.MAX_SITES; i++) { + baseline.record("site" + i, "host.com"); + } + // 触碰 site0，使其变为最近使用 + baseline.hosts("site0"); + baseline.record("overflow", "host.com"); + + assertEquals(SitePlaylistHostBaseline.MAX_SITES, baseline.siteCount()); + // site0 因刚被访问而存活，site1 成为最久未用被淘汰 + assertTrue(baseline.hosts("site0").contains("host.com")); + assertTrue(baseline.hosts("site1").isEmpty()); + } + + @Test + public void survivesRoundTripThroughStorage() { + SitePlaylistHostBaseline.MemoryStorage storage = new SitePlaylistHostBaseline.MemoryStorage(); + SitePlaylistHostBaseline first = new SitePlaylistHostBaseline(storage); + first.record("siteA", "a.com", "b.com"); + first.record("siteB", "c.com"); + + SitePlaylistHostBaseline restored = new SitePlaylistHostBaseline(storage); + + assertEquals(List.of("a.com", "b.com"), restored.hosts("siteA")); + assertEquals(List.of("c.com"), restored.hosts("siteB")); + } + + @Test + public void escapesSiteKeyContainingSeparators() { + SitePlaylistHostBaseline.MemoryStorage storage = new SitePlaylistHostBaseline.MemoryStorage(); + SitePlaylistHostBaseline first = new SitePlaylistHostBaseline(storage); + String awkward = "site;with|odd,chars%here"; + first.record(awkward, "a.com"); + + SitePlaylistHostBaseline restored = new SitePlaylistHostBaseline(storage); + + assertEquals(List.of("a.com"), restored.hosts(awkward)); + assertEquals(1, restored.siteCount()); + } + + @Test + public void ignoresBlankAndNullInput() { + SitePlaylistHostBaseline baseline = + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage()); + + baseline.record(null, "a.com"); + baseline.record(" ", "a.com"); + baseline.record("siteA", (String[]) null); + baseline.record("siteA", "", " "); + + assertEquals(0, baseline.siteCount()); + assertTrue(baseline.hosts("siteA").isEmpty()); + assertTrue(baseline.hosts(null).isEmpty()); + } + + @Test + public void toleratesCorruptStoredValue() { + SitePlaylistHostBaseline.MemoryStorage storage = new SitePlaylistHostBaseline.MemoryStorage(); + storage.write("garbage;;|;siteA|a.com;noseparator"); + + SitePlaylistHostBaseline baseline = new SitePlaylistHostBaseline(storage); + + assertEquals(List.of("a.com"), baseline.hosts("siteA")); + assertEquals(1, baseline.siteCount()); + } + + @Test + public void clearRemovesEverythingAndPersists() { + SitePlaylistHostBaseline.MemoryStorage storage = new SitePlaylistHostBaseline.MemoryStorage(); + SitePlaylistHostBaseline baseline = new SitePlaylistHostBaseline(storage); + baseline.record("siteA", "a.com"); + + baseline.clear(); + + assertEquals(0, baseline.siteCount()); + assertEquals(0, new SitePlaylistHostBaseline(storage).siteCount()); + } +} From d8ef2aed1f7fa6b6e6887af6d4175bc844b3bbd0 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sat, 29 Aug 2026 23:41:53 +0800 Subject: [PATCH 06/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E5=8C=BA?= =?UTF-8?q?=E9=97=B4=E8=B7=B3=E8=BF=87=E7=9A=84=E8=BF=90=E8=A1=8C=E6=97=B6?= =?UTF-8?q?=E6=8E=A5=E7=BA=BF=EF=BC=8C=E4=B8=8D=E4=BE=9D=E8=B5=96=E9=9F=B3?= =?UTF-8?q?=E9=A2=91=E6=8C=87=E7=BA=B9=E5=BC=80=E5=85=B3?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第六个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 9 节。 - AdAudioRuntimeController.skipUserInterval()：UI 已绑定但 coordinator 尚未 建立时按需创建。bindUi() 只在音频通道启用路径上被调用，不能作为区间跳过 的前置条件 —— 指纹与语音都关闭、规则库为空时，用户框选广告仍须能跳过。 - PlayerManager.skipUserAdInterval()：对外入口，isReleased() 时安全返回 false。 区间反馈单独跟踪 intervalSessionId/intervalGeneration 并在提交前对齐 coordinator：timeline reset 只经 PCM provider 的回调转发，音频通道关闭时 那条路径不存在，coordinator 会永远卡在上一次跳过留下的 UNDO_WINDOW 上导致 后续每次提交都被拒。基线取 seek 之后的实际状态，因为 seek 会推进 generation。 刻意不复用 activeSessionId：该字段决定音频管线是否需要重建，从区间路径改写 它会让 refreshLocked 漏掉一次必要的重建。 Verification: testMobileArm64_v8aDebugUnitTest --tests com.fongmi.android.tv.ad.* : 350 tests, 0 failures (AdAudioRuntimeControllerUserIntervalTest 8 new); compileMobileArm64_v8aDebugJavaWithJavac OK Task-Guard: AF4b-skip-plumbing --- .../tv/ad/audio/AdAudioRuntimeController.java | 45 ++++ .../android/tv/player/PlayerManager.java | 10 + ...udioRuntimeControllerUserIntervalTest.java | 232 ++++++++++++++++++ 3 files changed, 287 insertions(+) create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/audio/AdAudioRuntimeControllerUserIntervalTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/audio/AdAudioRuntimeController.java b/app/src/main/java/com/fongmi/android/tv/ad/audio/AdAudioRuntimeController.java index ff80383daa..11a61706f8 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/audio/AdAudioRuntimeController.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/audio/AdAudioRuntimeController.java @@ -61,6 +61,9 @@ public final class AdAudioRuntimeController implements AutoCloseable { private boolean enabled; private String lastRefreshLog = ""; private long activeSessionId = Long.MIN_VALUE; + /** 区间反馈单独跟踪 session/generation，避免与音频管线的重建判定耦合。 */ + private long intervalSessionId = Long.MIN_VALUE; + private long intervalGeneration = Long.MIN_VALUE; private boolean closed; public AdAudioRuntimeController(PlaybackMediaSignalHub hub, PlaybackMediaClock clock, @@ -172,12 +175,14 @@ public final class AdAudioRuntimeController implements AutoCloseable { if (coordinator != null) coordinator.close(); this.ui = ui; this.coordinator = new AdSkipCoordinator(playback, ui, 5_000L, diagnostics); + resetIntervalBaseline(); refreshLocked(); } public synchronized void unbindUi() { if (coordinator != null) coordinator.close(); coordinator = null; + resetIntervalBaseline(); ui = null; deactivateLocked(); } @@ -207,6 +212,39 @@ public final class AdAudioRuntimeController implements AutoCloseable { return snapshot; } + /** + * 执行用户框选区间的即时跳过。 + * + *

区间反馈不依赖音频指纹功能：即使指纹与语音通道都关闭、没有任何规则， + * 用户依然可以框选广告并要求跳过。因此这里在 UI 已绑定但 coordinator + * 尚未建立时按需创建一个 —— {@link #bindUi} 只在音频通道启用路径上被调用， + * 不能作为区间跳过的前置条件。 + * + * @return 是否真的执行了 seek；未绑定 UI 或校验失败时为 false + */ + public synchronized boolean skipUserInterval(long startMs, long endMs, String feedbackId) { + if (closed || ui == null) return false; + PlaybackMediaSignalHub.Session session = hub.session(); + if (coordinator == null) { + coordinator = new AdSkipCoordinator(playback, ui, 5_000L, diagnostics); + } else if (intervalSessionId != Long.MIN_VALUE + && (intervalSessionId != session.id() || intervalGeneration != session.generation())) { + // timeline reset 只经 PCM provider 的回调转发给 coordinator。音频通道关闭时 + // 那条路径不存在，coordinator 会永远卡在上一次跳过留下的 UNDO_WINDOW 上， + // 之后每一次区间提交都被拒。这里按播放实际状态主动对齐。 + // 不复用 activeSessionId：那个字段决定音频管线是否需要重建， + // 从这条路径改写它会让 refreshLocked 漏掉一次必要的重建。 + coordinator.reset(intervalSessionId); + } + boolean applied = coordinator.onUserInterval(session.id(), session.generation(), + startMs, endMs, feedbackId); + // seek 会推进 generation，基线取调用后的实际状态 + PlaybackMediaSignalHub.Session current = hub.session(); + intervalSessionId = current.id(); + intervalGeneration = current.generation(); + return applied; + } + public synchronized AdSkipPolicyController.Mode skipMode() { return skipMode; } @@ -236,6 +274,7 @@ public final class AdAudioRuntimeController implements AutoCloseable { deactivateLocked(); if (coordinator != null) coordinator.close(); coordinator = null; + resetIntervalBaseline(); ui = null; } @@ -246,10 +285,16 @@ public final class AdAudioRuntimeController implements AutoCloseable { deactivateLocked(); if (coordinator != null) coordinator.close(); coordinator = null; + resetIntervalBaseline(); ui = null; if (workerShutdown != null) workerShutdown.run(); } + private void resetIntervalBaseline() { + intervalSessionId = Long.MIN_VALUE; + intervalGeneration = Long.MIN_VALUE; + } + private void loadRulesLocked() { try { AdAudioRuleSnapshot loaded = ruleSource.load(); diff --git a/app/src/main/java/com/fongmi/android/tv/player/PlayerManager.java b/app/src/main/java/com/fongmi/android/tv/player/PlayerManager.java index f376bbd4e2..8efa519584 100644 --- a/app/src/main/java/com/fongmi/android/tv/player/PlayerManager.java +++ b/app/src/main/java/com/fongmi/android/tv/player/PlayerManager.java @@ -744,6 +744,16 @@ public class PlayerManager implements ParseCallback { adAudioRuntime.unbindUi(); } + /** + * 跳过用户框选的广告区间。区间为媒体时间，不依赖音频指纹功能开关。 + * + * @return 是否真的执行了 seek。false 表示已记录但未跳过（如终点已播过）。 + */ + public boolean skipUserAdInterval(long startMs, long endMs, String feedbackId) { + if (isReleased()) return false; + return adAudioRuntime.skipUserInterval(startMs, endMs, feedbackId); + } + public void reloadAdAudioRules() { reloadAdAudioSettings(); } diff --git a/app/src/test/java/com/fongmi/android/tv/ad/audio/AdAudioRuntimeControllerUserIntervalTest.java b/app/src/test/java/com/fongmi/android/tv/ad/audio/AdAudioRuntimeControllerUserIntervalTest.java new file mode 100644 index 0000000000..84fc8d83ae --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/audio/AdAudioRuntimeControllerUserIntervalTest.java @@ -0,0 +1,232 @@ +package com.fongmi.android.tv.ad.audio; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import com.fongmi.android.tv.player.audio.PlaybackMediaClock; +import com.fongmi.android.tv.player.audio.PlaybackMediaSignalHub; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +/** + * {@link AdAudioRuntimeController#skipUserInterval} 的行为锁定。 + * + *

核心契约：区间反馈不依赖音频指纹功能。指纹与语音都关闭、规则库为空时， + * 用户框选广告依然要能跳过。 + */ +public class AdAudioRuntimeControllerUserIntervalTest { + + @Test + public void skipsIntervalWithFingerprintDisabledAndNoRules() { + PlaybackMediaSignalHub hub = new PlaybackMediaSignalHub(8); + hub.beginSession(0L); + FakePlaybackPort playback = new FakePlaybackPort(hub, true); + AdAudioRuntimeController runtime = runtime(hub, playback, emptySnapshot()); + // 功能关闭、规则为空 —— 音频通道完全不工作 + runtime.start(false); + runtime.bindUi(new FakeUiPort()); + + assertFalse(hub.isCaptureRequested(PlaybackMediaSignalHub.ConsumerKind.AD_AUDIO)); + assertTrue(runtime.skipUserInterval(1_000L, 40_000L, "fb-1")); + assertEquals(List.of(40_000L), playback.seekTargets); + + runtime.close(); + } + + @Test + public void requiresBoundUiBecauseUndoPromptHasNowhereToGo() { + PlaybackMediaSignalHub hub = new PlaybackMediaSignalHub(8); + hub.beginSession(0L); + FakePlaybackPort playback = new FakePlaybackPort(hub, true); + AdAudioRuntimeController runtime = runtime(hub, playback, emptySnapshot()); + runtime.start(false); + + assertFalse(runtime.skipUserInterval(1_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + + runtime.close(); + } + + @Test + public void showsUndoPromptAfterSkip() { + PlaybackMediaSignalHub hub = new PlaybackMediaSignalHub(8); + hub.beginSession(0L); + FakePlaybackPort playback = new FakePlaybackPort(hub, true); + AdAudioRuntimeController runtime = runtime(hub, playback, emptySnapshot()); + FakeUiPort ui = new FakeUiPort(); + runtime.start(false); + runtime.bindUi(ui); + + runtime.skipUserInterval(1_000L, 40_000L, "fb-1"); + + assertEquals(1, ui.undoShows); + // 不经确认提示，用户框选本身即为授权 + assertEquals(0, ui.candidateShows); + + runtime.close(); + } + + @Test + public void reusesCoordinatorCreatedByBindUi() { + PlaybackMediaSignalHub hub = new PlaybackMediaSignalHub(8); + hub.beginSession(0L); + FakePlaybackPort playback = new FakePlaybackPort(hub, true); + // 指纹启用：bindUi 已建立 coordinator + AdAudioRuntimeController runtime = runtime(hub, playback, goodSnapshot()); + FakeUiPort ui = new FakeUiPort(); + runtime.start(true); + runtime.bindUi(ui); + + assertTrue(runtime.skipUserInterval(1_000L, 40_000L, "fb-1")); + assertEquals(List.of(40_000L), playback.seekTargets); + + runtime.close(); + } + + @Test + public void ineligiblePlaybackIsRejected() { + PlaybackMediaSignalHub hub = new PlaybackMediaSignalHub(8); + hub.beginSession(0L); + // eligible=false 会让快照变成 live 且不可 seek + FakePlaybackPort playback = new FakePlaybackPort(hub, false); + AdAudioRuntimeController runtime = runtime(hub, playback, emptySnapshot()); + runtime.start(false); + runtime.bindUi(new FakeUiPort()); + + assertFalse(runtime.skipUserInterval(1_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + + runtime.close(); + } + + @Test + public void closedRuntimeRejectsInterval() { + PlaybackMediaSignalHub hub = new PlaybackMediaSignalHub(8); + hub.beginSession(0L); + FakePlaybackPort playback = new FakePlaybackPort(hub, true); + AdAudioRuntimeController runtime = runtime(hub, playback, emptySnapshot()); + runtime.start(false); + runtime.bindUi(new FakeUiPort()); + runtime.close(); + + assertFalse(runtime.skipUserInterval(1_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + } + + @Test + public void unbindUiPreventsFurtherIntervalSkips() { + PlaybackMediaSignalHub hub = new PlaybackMediaSignalHub(8); + hub.beginSession(0L); + FakePlaybackPort playback = new FakePlaybackPort(hub, true); + AdAudioRuntimeController runtime = runtime(hub, playback, emptySnapshot()); + runtime.start(false); + runtime.bindUi(new FakeUiPort()); + runtime.unbindUi(); + + assertFalse(runtime.skipUserInterval(1_000L, 40_000L, "fb-1")); + assertTrue(playback.seekTargets.isEmpty()); + + runtime.close(); + } + + @Test + public void usesCurrentSessionSoStaleIntervalsCannotSeek() { + PlaybackMediaSignalHub hub = new PlaybackMediaSignalHub(8); + hub.beginSession(0L); + FakePlaybackPort playback = new FakePlaybackPort(hub, true); + AdAudioRuntimeController runtime = runtime(hub, playback, emptySnapshot()); + runtime.start(false); + runtime.bindUi(new FakeUiPort()); + + // 换源：session 变更后旧区间不应再套用到新内容上 + assertTrue(runtime.skipUserInterval(1_000L, 40_000L, "fb-1")); + hub.resetTimeline(0L, PlaybackMediaSignalHub.ResetReason.SOURCE_CHANGED); + + // 新 session 下同一 feedbackId 视为新的一次提交 + assertTrue(runtime.skipUserInterval(1_000L, 50_000L, "fb-1")); + assertEquals(List.of(40_000L, 50_000L), playback.seekTargets); + + runtime.close(); + } + + private static AdAudioRuntimeController runtime( + PlaybackMediaSignalHub hub, FakePlaybackPort playback, AdAudioRuleSnapshot snapshot) { + return new AdAudioRuntimeController( + hub, new PlaybackMediaClock(500L), () -> snapshot, playback, + Runnable::run, () -> { }); + } + + private static AdAudioRuleSnapshot emptySnapshot() { + return new AdAudioRuleSnapshot( + "test", "", AudioFingerprintRuleSet.empty(), List.of(), ""); + } + + private static AdAudioRuleSnapshot goodSnapshot() { + AudioFingerprintRuleSet rules = AudioFingerprintRuleCodec.fromJson("{" + + "\"schemaVersion\":2,\"algorithm\":{" + + "\"id\":\"spectral-sequence-v2\",\"sampleRate\":16000," + + "\"windowMs\":512,\"hopMs\":256,\"bandCount\":16}," + + "\"rules\":[{\"id\":\"ad\",\"durationMs\":10000," + + "\"anchorOffsetMs\":0,\"anchorDurationMs\":3000," + + "\"fingerprint\":[\"32f0007c\",\"35c100e0\",\"3b8b01c0\",\"d30a0380\"]}]}"); + return new AdAudioRuleSnapshot("test", "v1", rules, List.of(), ""); + } + + private static final class FakePlaybackPort implements AdAudioRuntimeController.PlaybackPort { + private final PlaybackMediaSignalHub hub; + private final List seekTargets = new ArrayList<>(); + private boolean eligible; + + FakePlaybackPort(PlaybackMediaSignalHub hub, boolean eligible) { + this.hub = hub; + this.eligible = eligible; + } + + @Override + public boolean isEligible(long sessionId, long generation) { + return eligible; + } + + @Override + public AdSkipCoordinator.PlaybackSnapshot snapshot(long sessionId, long generation) { + return new AdSkipCoordinator.PlaybackSnapshot( + sessionId, generation, 1_000L, 100_000L, eligible, !eligible, + new PlaybackMediaClock.Snapshot(generation, 0L, 50_000L, 1_000L, true, true)); + } + + @Override + public AdSkipCoordinator.SeekResult seekTo(long sessionId, long generation, long positionMs) { + PlaybackMediaSignalHub.Session session = hub.session(); + boolean applied = eligible && session.id() == sessionId + && session.generation() == generation; + if (applied) seekTargets.add(positionMs); + return new AdSkipCoordinator.SeekResult(applied, session.id(), session.generation()); + } + } + + private static final class FakeUiPort implements AdSkipCoordinator.UiPort { + private int candidateShows; + private int undoShows; + private AdSkipCoordinator.Actions actions; + + @Override + public void showCandidate(AdSkipCoordinator.Prompt prompt, AdSkipCoordinator.Actions actions) { + candidateShows++; + this.actions = actions; + } + + @Override + public void showUndo(AdSkipCoordinator.UndoPrompt prompt, AdSkipCoordinator.Actions actions) { + undoShows++; + this.actions = actions; + } + + @Override + public void dismiss(long sessionId) { + } + } +} From 2822ea1f9e78293f174d9023e1cc6cdd973b1e6c Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sat, 29 Aug 2026 23:51:45 +0800 Subject: [PATCH 07/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E5=8C=BA?= =?UTF-8?q?=E9=97=B4=E5=8F=8D=E9=A6=88=E8=AF=81=E6=8D=AE=E9=87=87=E9=9B=86?= =?UTF-8?q?=E4=B8=8E=E5=BD=92=E5=9B=A0=E7=BC=96=E6=8E=92?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Verification: testMobileArm64_v8aDebugUnitTest: ad.feedback 84 tests, 0 failures (AdEvidenceCollector 12, AdFeedbackController 13 new) Task-Guard: AF4c-controller --- .../tv/ad/feedback/AdEvidenceCollector.java | 170 ++++++++++ .../tv/ad/feedback/AdFeedbackController.java | 180 ++++++++++ .../ad/feedback/AdEvidenceCollectorTest.java | 187 ++++++++++ .../ad/feedback/AdFeedbackControllerTest.java | 321 ++++++++++++++++++ 4 files changed, 858 insertions(+) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollector.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollectorTest.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollector.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollector.java new file mode 100644 index 0000000000..6e069f6c40 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollector.java @@ -0,0 +1,170 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.fongmi.android.tv.bean.M3u8Evidence; + +import java.net.URI; +import java.util.ArrayList; +import java.util.LinkedHashSet; +import java.util.List; +import java.util.Locale; +import java.util.Set; + +/** + * 把 {@code M3u8Parser} 已产出的 playlist 证据 + 用户区间，组装成 + * {@link AdIntervalEvidence}。纯计算，不发起网络请求，不触碰 Android 类型。 + * + *

见设计文档第 6.1、7.1 节。 + */ +public final class AdEvidenceCollector { + + /** 区间外切片最多保留多少个作为对照组，避免长剧集证据膨胀。 */ + static final int MAX_OUTSIDE_SAMPLES = 60; + + private AdEvidenceCollector() { + } + + /** + * 播放上下文，全部由调用方从 Activity/PlayerManager 取出，保持本类可单测。 + * + * @param siteKey 站点标识 + * @param siteName 站点名称 + * @param vodName 剧名 + * @param flagName 线路名 + * @param episodeName 集名 + * @param playUrl 播放地址，仅用于取 host/path，不入证据 + * @param hls 是否 HLS + */ + public record Context(String siteKey, String siteName, String vodName, + String flagName, String episodeName, String playUrl, boolean hls) { + + public Context { + siteKey = siteKey == null ? "" : siteKey; + siteName = siteName == null ? "" : siteName; + vodName = vodName == null ? "" : vodName; + flagName = flagName == null ? "" : flagName; + episodeName = episodeName == null ? "" : episodeName; + playUrl = playUrl == null ? "" : playUrl; + } + } + + /** + * 组装区间证据。 + * + * @param evidence playlist 证据，可为 null（非 HLS 或抓取失败） + * @param blacklistedHosts 现有广告域名黑名单，用于填充 matchedExistingHosts + * @param legacyHeuristicActive 旧启发式引擎当次是否生效 + */ + public static AdIntervalEvidence collect(Context context, M3u8Evidence evidence, + long startMs, long endMs, StartOrigin origin, + List blacklistedHosts, + boolean legacyHeuristicActive) { + if (context == null) throw new IllegalArgumentException("context is required"); + String playlistHost = hostOf(context.playUrl()); + String urlPath = pathOf(context.playUrl()); + + List inside = new ArrayList<>(); + List outside = new ArrayList<>(); + boolean bounded = false; + + if (evidence != null && !evidence.isEmpty()) { + List durations = evidence.getDurations(); + List segments = evidence.getSegments(); + Set discontinuities = new LinkedHashSet<>(); + for (Integer index : evidence.getDiscontinuities()) { + if (index != null) discontinuities.add(index); + } + Set insideIndices = new LinkedHashSet<>( + AdIntervalMapper.insideIndices(durations, startMs, endMs)); + + int limit = Math.min(segments.size(), durations.size()); + for (int i = 0; i < limit; i++) { + SegmentFact fact = factOf(i, segments.get(i), durations.get(i), + discontinuities.contains(i)); + if (insideIndices.contains(i)) inside.add(fact); + else if (outside.size() < MAX_OUTSIDE_SAMPLES) outside.add(fact); + } + bounded = boundedByDiscontinuity(insideIndices, discontinuities, limit); + } + + boolean crossDomain = !inside.isEmpty() && !playlistHost.isEmpty() + && inside.stream().allMatch(fact -> + !fact.host().isEmpty() && !fact.hostEndsWith(playlistHost)); + + return new AdIntervalEvidence( + startMs, endMs, origin, + context.siteKey(), context.siteName(), context.vodName(), + context.flagName(), context.episodeName(), + playlistHost, urlPath, context.hls(), + inside, outside, bounded, crossDomain, + List.of(), legacyHeuristicActive, + matchedHosts(inside, blacklistedHosts), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + /** + * 区间首尾是否都贴着断点边界。首端要求区间第一个切片自身带断点， + * 尾端要求区间后的第一个切片带断点（即广告块结束、正片重新开始）。 + */ + private static boolean boundedByDiscontinuity(Set insideIndices, + Set discontinuities, int total) { + if (insideIndices.isEmpty() || discontinuities.isEmpty()) return false; + int first = Integer.MAX_VALUE; + int last = Integer.MIN_VALUE; + for (Integer index : insideIndices) { + first = Math.min(first, index); + last = Math.max(last, index); + } + boolean head = discontinuities.contains(first); + // 区间正好到 playlist 结尾时，没有「之后的切片」，视为尾端成立 + boolean tail = last + 1 >= total || discontinuities.contains(last + 1); + return head && tail; + } + + private static List matchedHosts(List inside, List blacklist) { + if (inside.isEmpty() || blacklist == null || blacklist.isEmpty()) return List.of(); + Set matched = new LinkedHashSet<>(); + for (SegmentFact fact : inside) { + if (fact.host().isEmpty()) continue; + String host = fact.host().toLowerCase(Locale.US); + for (String entry : blacklist) { + if (entry == null || entry.isBlank()) continue; + String target = entry.toLowerCase(Locale.US).trim(); + if (host.equals(target) || host.endsWith("." + target) || host.contains(target)) { + matched.add(fact.host()); + break; + } + } + } + return List.copyOf(matched); + } + + private static SegmentFact factOf(int index, String url, Float duration, boolean discontinuity) { + double seconds = duration == null ? 0d : duration; + return new SegmentFact(index, hostOf(url), pathOf(url), seconds, discontinuity); + } + + /** 取 host，解析失败返回空串。相对 URI 没有 host，属正常情况。 */ + static String hostOf(String url) { + if (url == null || url.isBlank()) return ""; + try { + String host = URI.create(url.trim()).getHost(); + return host == null ? "" : host; + } catch (RuntimeException e) { + return ""; + } + } + + /** 取去掉 query 的 path。相对 URI 直接返回其自身去参数部分。 */ + static String pathOf(String url) { + if (url == null || url.isBlank()) return ""; + String trimmed = url.trim(); + try { + String path = URI.create(trimmed).getPath(); + if (path != null && !path.isEmpty()) return path; + } catch (RuntimeException ignored) { + // 落到下面的字符串截断 + } + int query = trimmed.indexOf('?'); + return query < 0 ? trimmed : trimmed.substring(0, query); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java new file mode 100644 index 0000000000..98b160adcb --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java @@ -0,0 +1,180 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.fongmi.android.tv.bean.M3u8Evidence; + +import java.util.ArrayList; +import java.util.List; +import java.util.UUID; + +/** + * 区间反馈的编排器：接收区间 → 立即跳过 → 采集证据 → 多通道归因 → 仲裁 → 回调 UI。 + * + *

三端（leanback / mobile / TmdbDetail）共用本类，Activity 只实现 {@link Host}。 + * 见设计文档第 5.2 节。所有 Android 依赖都在 {@link Host} 后面，因此本类可单测。 + */ +public final class AdFeedbackController { + + /** Activity 侧需要提供的能力。 */ + public interface Host { + /** 当前播放位置，毫秒。 */ + long positionMs(); + + /** 总时长，毫秒；未知时返回非正数。 */ + long durationMs(); + + /** 采集上下文。 */ + AdEvidenceCollector.Context context(); + + /** 抓取 playlist 证据，阻塞 I/O，由 {@link #runBackground} 调度。可返回 null。 */ + M3u8Evidence fetchEvidence(); + + /** 执行本次跳过，返回是否真的 seek 了。 */ + boolean skipInterval(long startMs, long endMs, String feedbackId); + + /** 现有广告域名黑名单。 */ + List blacklistedHosts(); + + /** 本站历史正常域名。 */ + List siteBaselineHosts(); + + /** 接口学习的待审候选域名。 */ + List interfaceCandidateHosts(); + + /** 命中接口候选时展示的来源名。 */ + String interfaceSourceName(); + + /** 已知 HLS 规则状态。 */ + List hlsRuleStates(); + + /** 正片保护正则。 */ + List protectingExcludes(); + + /** 旧启发式引擎当次是否生效。 */ + boolean legacyHeuristicActive(); + + /** 音频/语音通道已有候选的起点，无则返回负数。 */ + default long audioCandidateStartMs() { + return -1L; + } + + /** 短按回溯窗口。 */ + default long fallbackWindowMs() { + return AdIntervalMapper.DEFAULT_FALLBACK_WINDOW_MS; + } + + /** 后台执行阻塞任务。 */ + void runBackground(Runnable task); + + /** 回到主线程更新 UI。 */ + void runOnUi(Runnable task); + + /** 展示或刷新反馈对话框。 */ + void showSession(AdFeedbackSession session); + } + + private final Host host; + private final AdFeedbackDiagnostics diagnostics; + /** 防止过期回调覆盖新一次反馈的结果。 */ + private int generation; + + public AdFeedbackController(Host host) { + this(host, new AdFeedbackDiagnostics()); + } + + public AdFeedbackController(Host host, AdFeedbackDiagnostics diagnostics) { + if (host == null) throw new IllegalArgumentException("host is required"); + this.host = host; + this.diagnostics = diagnostics == null ? new AdFeedbackDiagnostics() : diagnostics; + } + + public AdFeedbackDiagnostics diagnostics() { + return diagnostics; + } + + /** + * 短按：以当前位置为终点，回溯推断起点。 + * + * @return 本次反馈的会话，播放状态不允许时返回 null + */ + public AdFeedbackSession onQuickReport(M3u8Evidence cachedEvidence) { + long endMs = host.positionMs(); + if (endMs <= 0) return null; + AdIntervalMapper.InferredStart start = AdIntervalMapper.inferStart( + cachedEvidence, endMs, host.audioCandidateStartMs(), host.fallbackWindowMs()); + return submit(start.startMs(), endMs, start.origin()); + } + + /** 长按标记模式：用户显式框选的区间。 */ + public AdFeedbackSession onMarkedInterval(long startMs, long endMs) { + return submit(startMs, endMs, StartOrigin.USER_MARKED); + } + + /** + * 提交一个区间：立即尝试跳过，然后在后台归因。 + */ + public AdFeedbackSession submit(long startMs, long endMs, StartOrigin origin) { + if (endMs <= startMs || startMs < 0) return null; + long duration = host.durationMs(); + if (duration > 0 && startMs >= duration) return null; + + String feedbackId = "fb-" + UUID.randomUUID(); + diagnostics.record(AdFeedbackDiagnostics.Code.INTERVAL_SUBMITTED); + diagnostics.recordStartOrigin(origin); + + boolean skipped = host.skipInterval(startMs, endMs, feedbackId); + diagnostics.record(skipped + ? AdFeedbackDiagnostics.Code.IMMEDIATE_SKIP_APPLIED + : AdFeedbackDiagnostics.Code.IMMEDIATE_SKIP_REJECTED); + + AdFeedbackSession session = AdFeedbackSession.pending( + feedbackId, startMs, endMs, origin, skipped); + host.showSession(session); + + int current = ++generation; + host.runBackground(() -> analyse(session, current)); + return session; + } + + private void analyse(AdFeedbackSession session, int expectedGeneration) { + AdAttributionArbiter.Verdict verdict; + try { + verdict = classify(session); + } catch (RuntimeException e) { + diagnostics.record(AdFeedbackDiagnostics.Code.EVIDENCE_COLLECT_FAILED); + verdict = new AdAttributionArbiter.Verdict(null, List.of(), List.of()); + } + AdAttributionArbiter.Verdict result = verdict; + host.runOnUi(() -> { + if (expectedGeneration != generation) return; + host.showSession(session.withVerdict(result)); + }); + } + + private AdAttributionArbiter.Verdict classify(AdFeedbackSession session) { + M3u8Evidence playlist = host.fetchEvidence(); + List blacklist = host.blacklistedHosts(); + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + host.context(), playlist, session.startMs(), session.endMs(), + session.startOrigin(), blacklist, host.legacyHeuristicActive()); + + List attributions = new ArrayList<>(); + attributions.add(HlsSegmentClassifier.classify(evidence)); + attributions.add(DomainReputationClassifier.classify(evidence, + new DomainReputationClassifier.Input(blacklist, host.siteBaselineHosts(), + host.interfaceCandidateHosts(), host.interfaceSourceName()))); + attributions.add(ExistingRuleClassifier.classify(evidence, + new ExistingRuleClassifier.Input(host.hlsRuleStates(), + host.legacyHeuristicActive(), host.protectingExcludes()))); + + for (AdAttribution attribution : attributions) { + if (attribution == null) diagnostics.record(AdFeedbackDiagnostics.Code.CHANNEL_ABSTAINED); + else if (attribution.category() == AdCategory.ALREADY_HANDLED) { + diagnostics.record(AdFeedbackDiagnostics.Code.ALREADY_HANDLED_DETECTED); + } + } + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(attributions); + if (verdict.empty()) diagnostics.record(AdFeedbackDiagnostics.Code.ARBITER_NO_PLAN); + return verdict; + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollectorTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollectorTest.java new file mode 100644 index 0000000000..5b2dc48f30 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollectorTest.java @@ -0,0 +1,187 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import com.fongmi.android.tv.bean.M3u8Evidence; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +public class AdEvidenceCollectorTest { + + private static final AdEvidenceCollector.Context CONTEXT = new AdEvidenceCollector.Context( + "site", "站点", "剧名", "线路", "第 1 集", + "https://v.example.com/play/index.m3u8?token=secret", true); + + /** 10 个切片，每个 10s；下标 3-5 来自广告域名。 */ + private static M3u8Evidence tenSegments(List discontinuities) { + List segments = new ArrayList<>(); + List durations = new ArrayList<>(); + List switches = new ArrayList<>(); + for (int i = 0; i < 10; i++) { + boolean ad = i >= 3 && i <= 5; + segments.add(ad + ? "https://ad-cdn.other.com/creative/" + i + ".ts?sid=x" + : "https://v.example.com/seg/" + i + ".ts"); + durations.add(10f); + switches.add(ad); + } + return M3u8Evidence.create(segments, discontinuities, durations, switches); + } + + @Test + public void splitsSegmentsIntoInsideAndOutside() { + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, tenSegments(List.of()), 30_000, 60_000, + StartOrigin.USER_MARKED, List.of(), false); + + assertEquals(List.of(3, 4, 5), evidence.inside().stream().map(SegmentFact::index).toList()); + assertEquals(7, evidence.outside().size()); + assertTrue(evidence.hasSegmentEvidence()); + } + + @Test + public void stripsQueryFromPlaylistAndSegmentUrls() { + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, tenSegments(List.of()), 30_000, 60_000, + StartOrigin.USER_MARKED, List.of(), false); + + // 播放地址的 token 不得进入证据 + assertEquals("/play/index.m3u8", evidence.urlPath()); + assertEquals("v.example.com", evidence.playlistHost()); + assertEquals("/creative/3.ts", evidence.inside().get(0).path()); + assertTrue(evidence.inside().stream().noneMatch(fact -> fact.path().contains("sid="))); + } + + @Test + public void detectsCrossDomainWhenAllInsideSegmentsAreForeign() { + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, tenSegments(List.of()), 30_000, 60_000, + StartOrigin.USER_MARKED, List.of(), false); + + assertTrue(evidence.crossDomain()); + } + + @Test + public void noCrossDomainWhenIntervalMixesHosts() { + // 区间跨到正片切片上，不能整体判为跨域 + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, tenSegments(List.of()), 30_000, 80_000, + StartOrigin.USER_MARKED, List.of(), false); + + assertFalse(evidence.crossDomain()); + } + + @Test + public void boundedByDiscontinuityRequiresBothEnds() { + // 断点在 3（广告块开始）与 6（正片恢复） + AdIntervalEvidence both = AdEvidenceCollector.collect( + CONTEXT, tenSegments(List.of(3, 6)), 30_000, 60_000, + StartOrigin.USER_MARKED, List.of(), false); + assertTrue(both.boundedByDiscontinuity()); + + // 只有首端有断点 + AdIntervalEvidence headOnly = AdEvidenceCollector.collect( + CONTEXT, tenSegments(List.of(3)), 30_000, 60_000, + StartOrigin.USER_MARKED, List.of(), false); + assertFalse(headOnly.boundedByDiscontinuity()); + } + + @Test + public void intervalReachingPlaylistEndCountsTailAsBounded() { + // 区间到最后一个切片，没有「之后的切片」可带断点 + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, tenSegments(List.of(7)), 70_000, 100_000, + StartOrigin.USER_MARKED, List.of(), false); + + assertTrue(evidence.boundedByDiscontinuity()); + } + + @Test + public void recordsBlacklistMatches() { + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, tenSegments(List.of()), 30_000, 60_000, + StartOrigin.USER_MARKED, List.of("ad-cdn.other.com"), false); + + assertEquals(List.of("ad-cdn.other.com"), evidence.matchedExistingHosts()); + } + + @Test + public void capsOutsideSamplesForLongPlaylists() { + List segments = new ArrayList<>(); + List durations = new ArrayList<>(); + List switches = new ArrayList<>(); + for (int i = 0; i < 500; i++) { + segments.add("https://v.example.com/seg/" + i + ".ts"); + durations.add(10f); + switches.add(false); + } + M3u8Evidence long_ = M3u8Evidence.create(segments, List.of(), durations, switches); + + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, long_, 30_000, 60_000, StartOrigin.USER_MARKED, List.of(), false); + + assertEquals(3, evidence.inside().size()); + assertEquals(AdEvidenceCollector.MAX_OUTSIDE_SAMPLES, evidence.outside().size()); + } + + @Test + public void handlesMissingPlaylistEvidence() { + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, null, 30_000, 60_000, StartOrigin.FALLBACK_WINDOW, List.of(), true); + + assertTrue(evidence.inside().isEmpty()); + assertFalse(evidence.hasSegmentEvidence()); + assertFalse(evidence.crossDomain()); + assertTrue(evidence.legacyHeuristicActive()); + // 上下文仍然保留 + assertEquals("v.example.com", evidence.playlistHost()); + } + + @Test + public void handlesEmptyEvidenceAndNonHlsContext() { + AdEvidenceCollector.Context nonHls = new AdEvidenceCollector.Context( + "site", "站点", "剧名", "线路", "第 1 集", + "https://v.example.com/play/video.mp4", false); + + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + nonHls, M3u8Evidence.empty(), 30_000, 60_000, + StartOrigin.FALLBACK_WINDOW, List.of(), false); + + assertFalse(evidence.hls()); + assertFalse(evidence.hasSegmentEvidence()); + } + + @Test + public void relativeSegmentUriHasNoHostButKeepsPath() { + M3u8Evidence relative = M3u8Evidence.create( + List.of("seg/0.ts", "seg/1.ts", "seg/2.ts"), + List.of(), List.of(10f, 10f, 10f), List.of(false, false, false)); + + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, relative, 0, 10_000, StartOrigin.USER_MARKED, List.of(), false); + + assertEquals("", evidence.inside().get(0).host()); + assertEquals("seg/0.ts", evidence.inside().get(0).path()); + // 无 host 时不能判为跨域 + assertFalse(evidence.crossDomain()); + } + + @Test + public void toleratesDurationsShorterThanSegments() { + // durations 缺项时按较短的长度处理，不抛异常 + M3u8Evidence mismatched = M3u8Evidence.create( + List.of("a.ts", "b.ts", "c.ts"), + List.of(), List.of(10f, 10f), List.of(false, false)); + + AdIntervalEvidence evidence = AdEvidenceCollector.collect( + CONTEXT, mismatched, 0, 10_000, StartOrigin.USER_MARKED, List.of(), false); + + assertEquals(1, evidence.inside().size()); + assertEquals(1, evidence.outside().size()); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java new file mode 100644 index 0000000000..86785af554 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java @@ -0,0 +1,321 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertNotNull; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import com.fongmi.android.tv.bean.M3u8Evidence; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +public class AdFeedbackControllerTest { + + /** 10 个 10s 切片，下标 3-5 来自广告域名，断点在 3 与 6。 */ + private static M3u8Evidence adBlockPlaylist() { + List segments = new ArrayList<>(); + List durations = new ArrayList<>(); + List switches = new ArrayList<>(); + for (int i = 0; i < 10; i++) { + boolean ad = i >= 3 && i <= 5; + segments.add(ad ? "https://ad-cdn.other.com/ads/" + i + ".ts" + : "https://v.example.com/seg/" + i + ".ts"); + durations.add(ad ? 6.4f : 10f); + switches.add(ad); + } + return M3u8Evidence.create(segments, List.of(3, 6), durations, switches); + } + + @Test + public void showsPendingSessionBeforeAnalysisThenVerdict() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + AdFeedbackController controller = new AdFeedbackController(host); + + AdFeedbackSession session = controller.onMarkedInterval(30_000, 60_000); + + assertNotNull(session); + // 第一屏先出，第二屏带归因 + assertEquals(2, host.shownSessions.size()); + assertFalse(host.shownSessions.get(0).analysisComplete()); + assertTrue(host.shownSessions.get(1).analysisComplete()); + assertTrue(host.shownSessions.get(1).hasActionablePlan()); + } + + @Test + public void skipsImmediatelyAndRecordsOutcome() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + AdFeedbackController controller = new AdFeedbackController(host); + + AdFeedbackSession session = controller.onMarkedInterval(30_000, 60_000); + + assertTrue(session.skipApplied()); + assertEquals(List.of("30000-60000"), host.skipCalls); + assertEquals(1, controller.diagnostics() + .count(AdFeedbackDiagnostics.Code.IMMEDIATE_SKIP_APPLIED)); + } + + @Test + public void recordsRejectedSkipWithoutFailingTheFeedback() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + host.skipResult = false; + AdFeedbackController controller = new AdFeedbackController(host); + + AdFeedbackSession session = controller.onMarkedInterval(30_000, 60_000); + + // 未跳过不影响归因继续 + assertFalse(session.skipApplied()); + assertTrue(host.shownSessions.get(1).hasActionablePlan()); + assertEquals(1, controller.diagnostics() + .count(AdFeedbackDiagnostics.Code.IMMEDIATE_SKIP_REJECTED)); + } + + @Test + public void quickReportInfersStartFromNearestDiscontinuity() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + host.position = 60_000; + AdFeedbackController controller = new AdFeedbackController(host); + + AdFeedbackSession session = controller.onQuickReport(host.playlist); + + assertNotNull(session); + // 断点在下标 3 与 6。广告切片是 6.4s，故下标 6 起点为 + // 3×10 + 3×6.4 = 49.2s，是 60s 之前最近的那个断点。 + assertEquals(49_200L, session.startMs()); + assertEquals(StartOrigin.DISCONTINUITY, session.startOrigin()); + assertEquals(1, controller.diagnostics() + .count(AdFeedbackDiagnostics.Code.START_INFERRED_FROM_DISCONTINUITY)); + } + + @Test + public void quickReportPicksEarlierDiscontinuityWhenCloserOneIsAhead() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + // 位置落在广告块中间：只有下标 3 的断点在它之前 + host.position = 40_000; + AdFeedbackController controller = new AdFeedbackController(host); + + AdFeedbackSession session = controller.onQuickReport(host.playlist); + + assertEquals(30_000L, session.startMs()); + assertEquals(StartOrigin.DISCONTINUITY, session.startOrigin()); + } + + @Test + public void quickReportFallsBackToWindowWithoutPlaylist() { + FakeHost host = new FakeHost(); + host.position = 200_000; + AdFeedbackController controller = new AdFeedbackController(host); + + AdFeedbackSession session = controller.onQuickReport(null); + + assertEquals(200_000L - AdIntervalMapper.DEFAULT_FALLBACK_WINDOW_MS, session.startMs()); + assertEquals(StartOrigin.FALLBACK_WINDOW, session.startOrigin()); + assertEquals(1, controller.diagnostics() + .count(AdFeedbackDiagnostics.Code.START_INFERRED_FROM_FALLBACK)); + } + + @Test + public void rejectsInvalidIntervals() { + FakeHost host = new FakeHost(); + AdFeedbackController controller = new AdFeedbackController(host); + + assertNull(controller.onMarkedInterval(40_000, 40_000)); + assertNull(controller.onMarkedInterval(50_000, 40_000)); + assertNull(controller.onMarkedInterval(-1_000, 40_000)); + // 起点超出总时长 + assertNull(controller.onMarkedInterval(700_000, 800_000)); + assertTrue(host.skipCalls.isEmpty()); + assertTrue(host.shownSessions.isEmpty()); + } + + @Test + public void quickReportAtZeroPositionIsRejected() { + FakeHost host = new FakeHost(); + host.position = 0; + AdFeedbackController controller = new AdFeedbackController(host); + + assertNull(controller.onQuickReport(null)); + } + + @Test + public void surfacesAlreadyHandledDiagnosisWhenHostIsBlacklisted() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + host.blacklist = List.of("ad-cdn.other.com"); + AdFeedbackController controller = new AdFeedbackController(host); + + controller.onMarkedInterval(30_000, 60_000); + + AdFeedbackSession result = host.shownSessions.get(1); + assertTrue(result.verdict().diagnostics().stream() + .anyMatch(a -> a.category() == AdCategory.ALREADY_HANDLED)); + assertEquals(1, controller.diagnostics() + .count(AdFeedbackDiagnostics.Code.ALREADY_HANDLED_DETECTED)); + } + + @Test + public void prefersEnablingDisabledRuleOverNewRule() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + host.ruleStates = List.of(new ExistingRuleClassifier.RuleState( + "builtin:x", "x-rule", "实验规则", false, true, List.of("ad-cdn.other.com"))); + AdFeedbackController controller = new AdFeedbackController(host); + + controller.onMarkedInterval(30_000, 60_000); + + assertEquals(RemediationKind.ENABLE_EXISTING_RULE, + host.shownSessions.get(1).verdict().preferred().remediation()); + } + + @Test + public void fallsBackToSessionSkipWhenNoChannelHasEvidence() { + FakeHost host = new FakeHost(); + // 无 playlist、无基线、无规则：全部通道弃权 + host.legacyActive = true; + AdFeedbackController controller = new AdFeedbackController(host); + + controller.onMarkedInterval(30_000, 60_000); + + AdFeedbackSession result = host.shownSessions.get(1); + assertTrue(result.analysisComplete()); + assertFalse(result.hasActionablePlan()); + assertEquals(3, controller.diagnostics() + .count(AdFeedbackDiagnostics.Code.CHANNEL_ABSTAINED)); + } + + @Test + public void evidenceFailureStillCompletesWithEmptyVerdict() { + FakeHost host = new FakeHost(); + host.throwOnFetch = true; + AdFeedbackController controller = new AdFeedbackController(host); + + controller.onMarkedInterval(30_000, 60_000); + + AdFeedbackSession result = host.shownSessions.get(1); + assertTrue(result.analysisComplete()); + assertTrue(result.verdict().empty()); + assertEquals(1, controller.diagnostics() + .count(AdFeedbackDiagnostics.Code.EVIDENCE_COLLECT_FAILED)); + } + + @Test + public void staleAnalysisDoesNotOverwriteNewerFeedback() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + host.deferBackground = true; + AdFeedbackController controller = new AdFeedbackController(host); + + controller.onMarkedInterval(30_000, 60_000); + controller.onMarkedInterval(70_000, 90_000); + // 按倒序执行：先跑第二次的分析，再跑第一次的过期分析 + List pending = new ArrayList<>(host.backgroundTasks); + host.deferBackground = false; + pending.get(1).run(); + pending.get(0).run(); + + // 最后一次展示必须属于第二个区间 + AdFeedbackSession last = host.shownSessions.get(host.shownSessions.size() - 1); + assertEquals(70_000L, last.startMs()); + } + + private static final class FakeHost implements AdFeedbackController.Host { + private final List skipCalls = new ArrayList<>(); + private final List shownSessions = new ArrayList<>(); + private final List backgroundTasks = new ArrayList<>(); + private M3u8Evidence playlist; + private long position = 45_000; + private boolean skipResult = true; + private boolean legacyActive; + private boolean throwOnFetch; + private boolean deferBackground; + private List blacklist = List.of(); + private List ruleStates = List.of(); + + @Override + public long positionMs() { + return position; + } + + @Override + public long durationMs() { + return 600_000L; + } + + @Override + public AdEvidenceCollector.Context context() { + return new AdEvidenceCollector.Context("site", "站点", "剧名", "线路", "第 1 集", + "https://v.example.com/play/index.m3u8", true); + } + + @Override + public M3u8Evidence fetchEvidence() { + if (throwOnFetch) throw new IllegalStateException("fetch failed"); + return playlist; + } + + @Override + public boolean skipInterval(long startMs, long endMs, String feedbackId) { + skipCalls.add(startMs + "-" + endMs); + return skipResult; + } + + @Override + public List blacklistedHosts() { + return blacklist; + } + + @Override + public List siteBaselineHosts() { + return List.of("v.example.com"); + } + + @Override + public List interfaceCandidateHosts() { + return List.of(); + } + + @Override + public String interfaceSourceName() { + return ""; + } + + @Override + public List hlsRuleStates() { + return ruleStates; + } + + @Override + public List protectingExcludes() { + return List.of(); + } + + @Override + public boolean legacyHeuristicActive() { + return legacyActive; + } + + @Override + public void runBackground(Runnable task) { + backgroundTasks.add(task); + if (!deferBackground) task.run(); + } + + @Override + public void runOnUi(Runnable task) { + task.run(); + } + + @Override + public void showSession(AdFeedbackSession session) { + shownSessions.add(session); + } + } +} From d95c5537fced9da79cb055f62c322b0181e3e1fb Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sat, 29 Aug 2026 23:59:05 +0800 Subject: [PATCH 08/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E5=BD=92?= =?UTF-8?q?=E5=9B=A0=E6=95=B0=E6=8D=AE=E6=BA=90=E4=B8=8E=E7=AB=99=E7=82=B9?= =?UTF-8?q?=E5=9F=9F=E5=90=8D=E5=9F=BA=E7=BA=BF=E7=9A=84=E5=BA=94=E7=94=A8?= =?UTF-8?q?=E7=BA=A7=E8=90=BD=E5=9C=B0?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第八个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 5.1、8.3 节。 - AdFeedbackDataSource：把 RuleConfig / UserAdRuleStore / ImportedAdRuleCandidateStore / HlsRuleConfig 的静态单例数据读成分类器入参。 三个分类器刻意做成纯函数以便单测，读 SharedPreferences 集中在这里， 每个读取点都吞 RuntimeException —— 归因是辅助功能，不能因为读配置失败 影响播放。 - SiteHostBaselineStore：SitePlaylistHostBaseline 的 Prefers 实现与单例。 hostSuffixesOf 从 HlsRuleConfig.Entry.detail() 的 JSON 里取切片域名条件： HlsAdRule 的字段全部私有且无 getter，而 Entry.detail() 已经是它的完整 JSON 序列化，不必为此给上游 bean 加访问器。解析失败一律返回空列表。 尚未接线到播放器，无行为变化。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback: 89 tests, 0 failures (AdFeedbackDataSourceTest 5 new); compileMobileArm64_v8aDebug + compileLeanbackArm64_v8aDebug OK Task-Guard: AF4d-android-bridge --- .../tv/ad/feedback/AdFeedbackDataSource.java | 125 ++++++++++++++++++ .../tv/ad/feedback/SiteHostBaselineStore.java | 45 +++++++ .../ad/feedback/AdFeedbackDataSourceTest.java | 77 +++++++++++ 3 files changed, 247 insertions(+) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/SiteHostBaselineStore.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSourceTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java new file mode 100644 index 0000000000..15d864cb42 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java @@ -0,0 +1,125 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.fongmi.android.tv.api.config.HlsRuleConfig; +import com.fongmi.android.tv.api.config.ImportedAdRuleCandidateStore; +import com.fongmi.android.tv.api.config.RuleConfig; +import com.fongmi.android.tv.api.config.UserAdRuleStore; +import com.fongmi.android.tv.bean.ImportedAdRuleCandidate; +import com.fongmi.android.tv.bean.UserAdRule; +import com.google.gson.JsonArray; +import com.google.gson.JsonElement; +import com.google.gson.JsonObject; +import com.google.gson.JsonParser; + +import java.util.ArrayList; +import java.util.LinkedHashSet; +import java.util.List; +import java.util.Set; + +/** + * 把项目里的静态单例数据读成分类器需要的入参。 + * + *

三个分类器刻意做成纯函数以便单测，因此「读 SharedPreferences」这件事 + * 集中在这里。见设计文档第 5.1 节。 + */ +public final class AdFeedbackDataSource { + + private AdFeedbackDataSource() { + } + + /** 现有广告域名黑名单：VOD ads + Live ads + 用户规则 hosts。 */ + public static List blacklistedHosts() { + try { + return List.copyOf(RuleConfig.get().getAds()); + } catch (RuntimeException e) { + return List.of(); + } + } + + /** 待审接口候选里被判为广告域名的条目。 */ + public static List interfaceCandidateHosts() { + try { + Set hosts = new LinkedHashSet<>(); + for (ImportedAdRuleCandidate candidate : ImportedAdRuleCandidateStore.pending()) { + if (candidate == null) continue; + hosts.addAll(candidate.getHosts()); + } + return List.copyOf(hosts); + } catch (RuntimeException e) { + return List.of(); + } + } + + /** 第一条待审候选的来源接口名，用于证据展示。 */ + public static String interfaceSourceName() { + try { + for (ImportedAdRuleCandidate candidate : ImportedAdRuleCandidateStore.pending()) { + if (candidate == null) continue; + String name = candidate.getSourceConfigName(); + if (!name.isEmpty()) return name; + } + } catch (RuntimeException ignored) { + // 读取失败不影响归因 + } + return ""; + } + + /** + * 用户规则中的正片保护正则。这些规则可能误保护了广告切片， + * 是 {@link ExistingRuleClassifier} 的诊断输入之一。 + */ + public static List protectingExcludes() { + try { + List excludes = new ArrayList<>(); + for (UserAdRule rule : UserAdRuleStore.load()) { + if (rule == null || !rule.isEnabled()) continue; + excludes.addAll(rule.getExclude()); + } + return List.copyOf(excludes); + } catch (RuntimeException e) { + return List.of(); + } + } + + /** + * 已知 HLS 规则状态。{@code HlsAdRule} 未暴露 hostSuffixes 的 getter， + * 因此从 {@link HlsRuleConfig.Entry#detail()} 的 JSON 里取。 + */ + public static List hlsRuleStates() { + try { + List states = new ArrayList<>(); + for (HlsRuleConfig.Entry entry : HlsRuleConfig.getEntries()) { + if (entry == null) continue; + states.add(new ExistingRuleClassifier.RuleState( + entry.key(), entry.id(), entry.name(), + entry.enabled(), entry.valid(), + hostSuffixesOf(entry.detail()))); + } + return List.copyOf(states); + } catch (RuntimeException e) { + return List.of(); + } + } + + /** 从规则详情 JSON 里解析 hostSuffixes。解析失败返回空列表，不抛异常。 */ + static List hostSuffixesOf(String detailJson) { + if (detailJson == null || detailJson.isBlank()) return List.of(); + try { + JsonElement parsed = JsonParser.parseString(detailJson); + if (!parsed.isJsonObject()) return List.of(); + JsonObject object = parsed.getAsJsonObject(); + JsonElement hosts = object.get("hostSuffixes"); + if (hosts == null || !hosts.isJsonArray()) return List.of(); + JsonArray array = hosts.getAsJsonArray(); + List result = new ArrayList<>(array.size()); + for (JsonElement element : array) { + if (element == null || element.isJsonNull()) continue; + String value = element.getAsString(); + if (value != null && !value.isBlank()) result.add(value.trim()); + } + return List.copyOf(result); + } catch (RuntimeException e) { + return List.of(); + } + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SiteHostBaselineStore.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SiteHostBaselineStore.java new file mode 100644 index 0000000000..b3470cdac8 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SiteHostBaselineStore.java @@ -0,0 +1,45 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.github.catvod.utils.Prefers; + +/** + * {@link SitePlaylistHostBaseline} 的应用级单例，持久化到 SharedPreferences。 + */ +public final class SiteHostBaselineStore { + + private static final String PREF_KEY = "ad_feedback_site_hosts"; + + private static volatile SitePlaylistHostBaseline instance; + + private SiteHostBaselineStore() { + } + + public static SitePlaylistHostBaseline get() { + if (instance == null) { + synchronized (SiteHostBaselineStore.class) { + if (instance == null) instance = new SitePlaylistHostBaseline(new PrefersStorage()); + } + } + return instance; + } + + private static final class PrefersStorage implements SitePlaylistHostBaseline.Storage { + @Override + public String read() { + try { + return Prefers.getString(PREF_KEY, ""); + } catch (RuntimeException e) { + return ""; + } + } + + @Override + public void write(String value) { + try { + Prefers.put(PREF_KEY, value == null ? "" : value); + } catch (RuntimeException ignored) { + // 基线只是归因的辅助输入，写失败不应影响播放 + } + } + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSourceTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSourceTest.java new file mode 100644 index 0000000000..7e750d33f3 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSourceTest.java @@ -0,0 +1,77 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.List; + +/** + * {@link AdFeedbackDataSource} 的纯解析部分。 + * + *

读取静态单例的方法依赖 Android 运行时，无法在此覆盖；这里只锁定 + * {@code hostSuffixesOf} 的解析行为，它是规则状态能否被正确识别的关键。 + */ +public class AdFeedbackDataSourceTest { + + @Test + public void parsesHostSuffixesFromRuleDetail() { + String detail = "{\"id\":\"baofeng\",\"hostSuffixes\":[\"ad-cdn.a.com\",\"ads.b.com\"]," + + "\"minimumSignals\":2}"; + + assertEquals(List.of("ad-cdn.a.com", "ads.b.com"), + AdFeedbackDataSource.hostSuffixesOf(detail)); + } + + @Test + public void trimsAndSkipsBlankEntries() { + String detail = "{\"hostSuffixes\":[\" ad.a.com \",\"\",\" \",null,\"ad.b.com\"]}"; + + assertEquals(List.of("ad.a.com", "ad.b.com"), + AdFeedbackDataSource.hostSuffixesOf(detail)); + } + + @Test + public void returnsEmptyWhenFieldMissingOrWrongType() { + assertTrue(AdFeedbackDataSource.hostSuffixesOf("{\"id\":\"x\"}").isEmpty()); + // 规则只用 playlistHostSuffixes 限定作用域，没有切片域名条件 + assertTrue(AdFeedbackDataSource.hostSuffixesOf( + "{\"playlistHostSuffixes\":[\"v.a.com\"]}").isEmpty()); + assertTrue(AdFeedbackDataSource.hostSuffixesOf("{\"hostSuffixes\":\"ad.a.com\"}").isEmpty()); + assertTrue(AdFeedbackDataSource.hostSuffixesOf("{\"hostSuffixes\":null}").isEmpty()); + } + + @Test + public void toleratesMalformedInput() { + assertTrue(AdFeedbackDataSource.hostSuffixesOf(null).isEmpty()); + assertTrue(AdFeedbackDataSource.hostSuffixesOf("").isEmpty()); + assertTrue(AdFeedbackDataSource.hostSuffixesOf(" ").isEmpty()); + assertTrue(AdFeedbackDataSource.hostSuffixesOf("not json").isEmpty()); + assertTrue(AdFeedbackDataSource.hostSuffixesOf("[1,2,3]").isEmpty()); + assertTrue(AdFeedbackDataSource.hostSuffixesOf("{\"hostSuffixes\":[").isEmpty()); + } + + @Test + public void parsedSuffixesFeedRuleStateMatching() { + // 端到端：解析出的后缀能让 ExistingRuleClassifier 命中区间切片 + List suffixes = AdFeedbackDataSource.hostSuffixesOf( + "{\"hostSuffixes\":[\"ad-cdn.other.com\"]}"); + ExistingRuleClassifier.RuleState disabled = new ExistingRuleClassifier.RuleState( + "builtin:x", "x", "实验规则", false, true, suffixes); + + AdIntervalEvidence evidence = new AdIntervalEvidence( + 10_000, 40_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + "v.example.com", "/play/index.m3u8", true, + List.of(new SegmentFact(3, "ad-cdn.other.com", "/seg/3.ts", 6.4, true)), + List.of(new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false)), + false, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + AdAttribution attribution = ExistingRuleClassifier.classify(evidence, + new ExistingRuleClassifier.Input(List.of(disabled), false, List.of())); + + assertEquals(RemediationKind.ENABLE_EXISTING_RULE, attribution.remediation()); + } +} From 48731cbbbf094b5178c2e7672a497ebb418763bc Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 00:08:24 +0800 Subject: [PATCH 09/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E4=B8=89?= =?UTF-8?q?=E7=AB=AF=E5=85=B1=E7=94=A8=E7=9A=84=E5=8F=8D=E9=A6=88=20Host?= =?UTF-8?q?=20=E9=80=82=E9=85=8D=E5=99=A8?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第九个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 5.2 节。 AdFeedbackHostAdapter 实现 AdFeedbackController.Host，把 Activity 需要提供的 东西收窄成三个小接口：Playback（播放状态与跳过）、Metadata（剧集元数据）、 Ui（线程调度与展示）。读本地配置、抓 playlist、维护站点域名基线都在适配器内， 三端 Activity 因此不必各写一份。 - 配置读取全部经 Deps 注入，默认指向 AdFeedbackDataSource，测试可替换； 每个读取点吞 RuntimeException 并降级为空集合。 - fetchEvidence 缓存最近一次成功抓取的证据供短按推断起点复用， invalidateEvidence() 供换源换集时清理，避免用上一集的切片结构。 - legacyHeuristicActive 对点播 HLS 恒为 true：HlsAdblockPipeline 两个调用点 都传 legacyFallback=true 且内置规则默认全关，旧启发式实际总在生效。 尚未接线到播放器，无行为变化。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback: 99 tests, 0 failures (AdFeedbackHostAdapterTest 10 new); compileMobileArm64_v8aDebug + compileLeanbackArm64_v8aDebug OK Task-Guard: AF4e-host-adapter --- .../tv/ad/feedback/AdFeedbackHostAdapter.java | 233 +++++++++++++++++ .../feedback/AdFeedbackHostAdapterTest.java | 237 ++++++++++++++++++ 2 files changed, 470 insertions(+) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapterTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java new file mode 100644 index 0000000000..7fd1723534 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java @@ -0,0 +1,233 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.fongmi.android.tv.bean.M3u8Evidence; +import com.fongmi.android.tv.utils.M3u8Parser; + +import java.util.List; +import java.util.Map; +import java.util.Objects; +import java.util.function.Supplier; + +/** + * {@link AdFeedbackController.Host} 的通用实现，三端共用。 + * + *

Activity 只需提供播放状态、剧集元数据和三个调度/展示回调；读取本地配置、 + * 抓取 playlist、维护站点域名基线都在这里完成。见设计文档第 5.2 节。 + */ +public final class AdFeedbackHostAdapter implements AdFeedbackController.Host { + + /** 播放器状态。{@code hls} 由调用方判定，避免本类依赖 media3。 */ + public interface Playback { + long positionMs(); + + long durationMs(); + + String playUrl(); + + Map headers(); + + boolean hls(); + + /** 执行本次跳过，通常委托给 {@code PlayerManager.skipUserAdInterval}。 */ + boolean skipInterval(long startMs, long endMs, String feedbackId); + + /** 音频/语音通道已有候选的起点，无则返回负数。 */ + default long audioCandidateStartMs() { + return -1L; + } + } + + /** 当前剧集元数据。 */ + public interface Metadata { + String siteKey(); + + String siteName(); + + String vodName(); + + String flagName(); + + String episodeName(); + } + + /** 线程调度与界面展示。 */ + public interface Ui { + void runBackground(Runnable task); + + void runOnUi(Runnable task); + + void showSession(AdFeedbackSession session); + } + + /** + * 本地配置读取入口，默认指向 {@link AdFeedbackDataSource}，测试可替换。 + */ + public record Deps(Supplier> blacklistedHosts, + Supplier> interfaceCandidateHosts, + Supplier interfaceSourceName, + Supplier> protectingExcludes, + Supplier> hlsRuleStates, + SitePlaylistHostBaseline baseline) { + + public static Deps defaults() { + return new Deps( + AdFeedbackDataSource::blacklistedHosts, + AdFeedbackDataSource::interfaceCandidateHosts, + AdFeedbackDataSource::interfaceSourceName, + AdFeedbackDataSource::protectingExcludes, + AdFeedbackDataSource::hlsRuleStates, + SiteHostBaselineStore.get()); + } + } + + private final Playback playback; + private final Metadata metadata; + private final Ui ui; + private final Deps deps; + /** 最近一次抓到的 playlist 证据，短按推断起点时复用，避免重复请求。 */ + private volatile M3u8Evidence cachedEvidence; + + public AdFeedbackHostAdapter(Playback playback, Metadata metadata, Ui ui) { + this(playback, metadata, ui, Deps.defaults()); + } + + public AdFeedbackHostAdapter(Playback playback, Metadata metadata, Ui ui, Deps deps) { + this.playback = Objects.requireNonNull(playback, "playback"); + this.metadata = Objects.requireNonNull(metadata, "metadata"); + this.ui = Objects.requireNonNull(ui, "ui"); + this.deps = deps == null ? Deps.defaults() : deps; + } + + /** 供短按路径复用的缓存证据，可能为 null。 */ + public M3u8Evidence cachedEvidence() { + return cachedEvidence; + } + + /** 换源或换集时清掉缓存，避免把上一集的切片结构用于本集。 */ + public void invalidateEvidence() { + cachedEvidence = null; + } + + /** 记录一次成功播放的域名，喂给站点基线。 */ + public void recordPlaybackHost() { + String host = AdEvidenceCollector.hostOf(playback.playUrl()); + if (host.isEmpty()) return; + deps.baseline().record(metadata.siteKey(), host); + } + + @Override + public long positionMs() { + return playback.positionMs(); + } + + @Override + public long durationMs() { + return playback.durationMs(); + } + + @Override + public AdEvidenceCollector.Context context() { + return new AdEvidenceCollector.Context( + metadata.siteKey(), metadata.siteName(), metadata.vodName(), + metadata.flagName(), metadata.episodeName(), playback.playUrl(), playback.hls()); + } + + @Override + public M3u8Evidence fetchEvidence() { + if (!playback.hls()) return null; + String url = playback.playUrl(); + if (url == null || url.isBlank()) return null; + try { + M3u8Evidence evidence = M3u8Parser.parse(url, playback.headers()); + if (evidence != null && !evidence.isEmpty()) cachedEvidence = evidence; + return evidence; + } catch (RuntimeException e) { + return null; + } + } + + @Override + public boolean skipInterval(long startMs, long endMs, String feedbackId) { + return playback.skipInterval(startMs, endMs, feedbackId); + } + + @Override + public List blacklistedHosts() { + return orEmpty(deps.blacklistedHosts()); + } + + @Override + public List siteBaselineHosts() { + try { + return deps.baseline().hosts(metadata.siteKey()); + } catch (RuntimeException e) { + return List.of(); + } + } + + @Override + public List interfaceCandidateHosts() { + return orEmpty(deps.interfaceCandidateHosts()); + } + + @Override + public String interfaceSourceName() { + try { + String name = deps.interfaceSourceName().get(); + return name == null ? "" : name; + } catch (RuntimeException e) { + return ""; + } + } + + @Override + public List hlsRuleStates() { + try { + List states = deps.hlsRuleStates().get(); + return states == null ? List.of() : states; + } catch (RuntimeException e) { + return List.of(); + } + } + + @Override + public List protectingExcludes() { + return orEmpty(deps.protectingExcludes()); + } + + @Override + public boolean legacyHeuristicActive() { + // HlsAdblockPipeline 在两个调用点都传 legacyFallback=true， + // 且内置规则默认全部关闭，所以点播 HLS 上旧启发式实际总是生效。 + return playback.hls(); + } + + @Override + public long audioCandidateStartMs() { + return playback.audioCandidateStartMs(); + } + + @Override + public void runBackground(Runnable task) { + ui.runBackground(task); + } + + @Override + public void runOnUi(Runnable task) { + ui.runOnUi(task); + } + + @Override + public void showSession(AdFeedbackSession session) { + ui.showSession(session); + } + + private static List orEmpty(Supplier> supplier) { + try { + List value = supplier.get(); + return value == null ? List.of() : value; + } catch (RuntimeException e) { + return List.of(); + } + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapterTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapterTest.java new file mode 100644 index 0000000000..2d7e838a03 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapterTest.java @@ -0,0 +1,237 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.List; +import java.util.Map; + +/** + * {@link AdFeedbackHostAdapter} 的可测部分：配置读取的容错、站点基线记录、 + * 以及不依赖网络的字段映射。{@code fetchEvidence} 会发起真实请求，此处不覆盖。 + */ +public class AdFeedbackHostAdapterTest { + + @Test + public void mapsMetadataAndPlaybackIntoContext() { + FakePlayback playback = new FakePlayback(); + AdFeedbackHostAdapter adapter = adapter(playback, deps()); + + AdEvidenceCollector.Context context = adapter.context(); + + assertEquals("site", context.siteKey()); + assertEquals("剧名", context.vodName()); + assertEquals("https://v.example.com/play/index.m3u8", context.playUrl()); + assertTrue(context.hls()); + assertEquals(45_000L, adapter.positionMs()); + assertEquals(600_000L, adapter.durationMs()); + } + + @Test + public void recordsPlaybackHostIntoBaseline() { + SitePlaylistHostBaseline baseline = + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage()); + AdFeedbackHostAdapter adapter = adapter(new FakePlayback(), deps(baseline)); + + adapter.recordPlaybackHost(); + + assertEquals(List.of("v.example.com"), baseline.hosts("site")); + assertEquals(List.of("v.example.com"), adapter.siteBaselineHosts()); + } + + @Test + public void skipsBaselineRecordWhenUrlHasNoHost() { + SitePlaylistHostBaseline baseline = + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage()); + FakePlayback playback = new FakePlayback(); + playback.url = "relative/path.m3u8"; + AdFeedbackHostAdapter adapter = adapter(playback, deps(baseline)); + + adapter.recordPlaybackHost(); + + assertEquals(0, baseline.siteCount()); + } + + @Test + public void configReadFailuresDegradeToEmptyInsteadOfCrashing() { + AdFeedbackHostAdapter.Deps throwing = new AdFeedbackHostAdapter.Deps( + () -> { throw new IllegalStateException("prefs unavailable"); }, + () -> { throw new IllegalStateException("prefs unavailable"); }, + () -> { throw new IllegalStateException("prefs unavailable"); }, + () -> { throw new IllegalStateException("prefs unavailable"); }, + () -> { throw new IllegalStateException("prefs unavailable"); }, + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage())); + AdFeedbackHostAdapter adapter = adapter(new FakePlayback(), throwing); + + // 归因是辅助功能，读配置失败必须降级而不是把异常抛给播放中的用户 + assertTrue(adapter.blacklistedHosts().isEmpty()); + assertTrue(adapter.interfaceCandidateHosts().isEmpty()); + assertEquals("", adapter.interfaceSourceName()); + assertTrue(adapter.protectingExcludes().isEmpty()); + assertTrue(adapter.hlsRuleStates().isEmpty()); + } + + @Test + public void nullConfigValuesBecomeEmptyLists() { + AdFeedbackHostAdapter.Deps nulls = new AdFeedbackHostAdapter.Deps( + () -> null, () -> null, () -> null, () -> null, () -> null, + new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage())); + AdFeedbackHostAdapter adapter = adapter(new FakePlayback(), nulls); + + assertTrue(adapter.blacklistedHosts().isEmpty()); + assertEquals("", adapter.interfaceSourceName()); + assertTrue(adapter.hlsRuleStates().isEmpty()); + } + + @Test + public void nonHlsPlaybackSkipsEvidenceFetchAndLegacyFlag() { + FakePlayback playback = new FakePlayback(); + playback.hls = false; + AdFeedbackHostAdapter adapter = adapter(playback, deps()); + + // 不是 HLS 就不发请求 + assertNull(adapter.fetchEvidence()); + // 旧启发式只在点播 HLS 上生效 + assertFalse(adapter.legacyHeuristicActive()); + } + + @Test + public void legacyHeuristicIsActiveForHlsPlayback() { + AdFeedbackHostAdapter adapter = adapter(new FakePlayback(), deps()); + + assertTrue(adapter.legacyHeuristicActive()); + } + + @Test + public void blankUrlSkipsEvidenceFetch() { + FakePlayback playback = new FakePlayback(); + playback.url = " "; + AdFeedbackHostAdapter adapter = adapter(playback, deps()); + + assertNull(adapter.fetchEvidence()); + } + + @Test + public void invalidateEvidenceClearsCache() { + AdFeedbackHostAdapter adapter = adapter(new FakePlayback(), deps()); + + adapter.invalidateEvidence(); + + assertNull(adapter.cachedEvidence()); + } + + @Test + public void delegatesSkipAndAudioCandidateToPlayback() { + FakePlayback playback = new FakePlayback(); + playback.audioStart = 12_000L; + AdFeedbackHostAdapter adapter = adapter(playback, deps()); + + assertTrue(adapter.skipInterval(10_000, 40_000, "fb-1")); + assertEquals(List.of("10000-40000/fb-1"), playback.skipCalls); + assertEquals(12_000L, adapter.audioCandidateStartMs()); + } + + private static AdFeedbackHostAdapter adapter(FakePlayback playback, + AdFeedbackHostAdapter.Deps deps) { + return new AdFeedbackHostAdapter(playback, new FakeMetadata(), new FakeUi(), deps); + } + + private static AdFeedbackHostAdapter.Deps deps() { + return deps(new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage())); + } + + private static AdFeedbackHostAdapter.Deps deps(SitePlaylistHostBaseline baseline) { + return new AdFeedbackHostAdapter.Deps( + List::of, List::of, () -> "", List::of, List::of, baseline); + } + + private static final class FakePlayback implements AdFeedbackHostAdapter.Playback { + private final List skipCalls = new java.util.ArrayList<>(); + private String url = "https://v.example.com/play/index.m3u8"; + private boolean hls = true; + private long audioStart = -1L; + + @Override + public long positionMs() { + return 45_000L; + } + + @Override + public long durationMs() { + return 600_000L; + } + + @Override + public String playUrl() { + return url; + } + + @Override + public Map headers() { + return Map.of(); + } + + @Override + public boolean hls() { + return hls; + } + + @Override + public boolean skipInterval(long startMs, long endMs, String feedbackId) { + skipCalls.add(startMs + "-" + endMs + "/" + feedbackId); + return true; + } + + @Override + public long audioCandidateStartMs() { + return audioStart; + } + } + + private static final class FakeMetadata implements AdFeedbackHostAdapter.Metadata { + @Override + public String siteKey() { + return "site"; + } + + @Override + public String siteName() { + return "站点"; + } + + @Override + public String vodName() { + return "剧名"; + } + + @Override + public String flagName() { + return "线路"; + } + + @Override + public String episodeName() { + return "第 1 集"; + } + } + + private static final class FakeUi implements AdFeedbackHostAdapter.Ui { + @Override + public void runBackground(Runnable task) { + task.run(); + } + + @Override + public void runOnUi(Runnable task) { + task.run(); + } + + @Override + public void showSession(AdFeedbackSession session) { + } + } +} From f19d4a0e97c1deb7fa764a6cd7db0ab77af89489 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 00:27:46 +0800 Subject: [PATCH 10/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E5=8C=BA?= =?UTF-8?q?=E9=97=B4=E5=8F=8D=E9=A6=88=E4=B8=A4=E5=B1=8F=E5=AF=B9=E8=AF=9D?= =?UTF-8?q?=E6=A1=86=E4=B8=8E=E4=B8=89=E8=AF=AD=E8=A8=80=E6=96=87=E6=A1=88?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第十个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 4.3 节。 AdFeedbackDialog 放在 main 源集三端共用，参照 AdSkipPromptPresenter 的模式： 用 MaterialAlertDialogBuilder + 代码搭建视图，不写 layout xml。理由是内容 本身是不定长的证据列表，用 layout 需要为 mobile/leanback 各维护一份文件， 而现有 dialog_ad_rule_preview.xml 已经是那样重复了两份。 - 第一屏提交瞬间即出：区间、时长、起点来源、是否已跳过。 - 第二屏归因完成后原地替换正文，同一 feedbackId 不重建对话框， 避免 TV 上闪一下丢焦点；只有「保存规则」按钮的有无发生变化时才重建， 因为 AlertDialog 不支持事后增删按钮。 - 起点来源逐一展示，让用户知道 02:15 是框选的还是推断的。 - 按钮在 TV 上居中，遥控器更容易命中。 新增 27 条文案，en/zh-rCN/zh-rTW 三份齐全，格式串统一用带位置索引的 %1$ 形式以通过 lint 的 StringFormat 检查。 对话框尚无调用方，无行为变化。 Verification: compileMobileArm64_v8aDebug + compileLeanbackArm64_v8aDebug OK; lintVitalMobileArm64_v8aRelease OK (no StringFormat/MissingTranslation) Task-Guard: AF4f-feedback-dialog --- .../tv/ui/dialog/AdFeedbackDialog.java | 246 ++++++++++++++++++ app/src/main/res/values-zh-rCN/strings.xml | 25 ++ app/src/main/res/values-zh-rTW/strings.xml | 25 ++ app/src/main/res/values/strings.xml | 25 ++ 4 files changed, 321 insertions(+) create mode 100644 app/src/main/java/com/fongmi/android/tv/ui/dialog/AdFeedbackDialog.java diff --git a/app/src/main/java/com/fongmi/android/tv/ui/dialog/AdFeedbackDialog.java b/app/src/main/java/com/fongmi/android/tv/ui/dialog/AdFeedbackDialog.java new file mode 100644 index 0000000000..82297d1f97 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ui/dialog/AdFeedbackDialog.java @@ -0,0 +1,246 @@ +package com.fongmi.android.tv.ui.dialog; + +import android.os.Handler; +import android.os.Looper; +import android.view.Gravity; +import android.view.View; +import android.widget.LinearLayout; +import android.widget.ScrollView; +import android.widget.TextView; + +import androidx.appcompat.app.AlertDialog; +import androidx.fragment.app.FragmentActivity; + +import com.fongmi.android.tv.R; +import com.fongmi.android.tv.ad.feedback.AdAttribution; +import com.fongmi.android.tv.ad.feedback.AdFeedbackSession; +import com.fongmi.android.tv.ad.feedback.RemediationKind; +import com.fongmi.android.tv.ad.feedback.RiskLevel; +import com.fongmi.android.tv.ad.feedback.StartOrigin; +import com.fongmi.android.tv.utils.ResUtil; +import com.google.android.material.dialog.MaterialAlertDialogBuilder; + +import java.lang.ref.WeakReference; +import java.util.List; +import java.util.Locale; +import java.util.concurrent.TimeUnit; + +/** + * 区间反馈对话框。第一屏在提交瞬间就出，第二屏在归因完成后原地替换内容。 + * + *

视图用代码搭建而非 layout：三端共用一份，且内容是不定长的证据列表， + * 用 layout 反而要为每端各维护一份文件。见设计文档第 4.3 节。 + */ +public final class AdFeedbackDialog implements AutoCloseable { + + /** 用户确认保存规则。 */ + public interface Callback { + void onSave(AdAttribution plan); + } + + private final WeakReference activity; + private final Handler main = new Handler(Looper.getMainLooper()); + private final Callback callback; + private AlertDialog dialog; + private TextView body; + private String feedbackId; + private boolean closed; + + public AdFeedbackDialog(FragmentActivity activity, Callback callback) { + this.activity = new WeakReference<>(activity); + this.callback = callback; + } + + /** 展示或原地刷新。同一 feedbackId 的后续调用只替换正文。 */ + public void show(AdFeedbackSession session) { + if (session == null) return; + runOnMain(() -> showInternal(session)); + } + + @Override + public void close() { + runOnMain(() -> { + if (closed) return; + closed = true; + dismissInternal(); + activity.clear(); + }); + } + + private void showInternal(AdFeedbackSession session) { + FragmentActivity owner = usableActivity(); + if (owner == null || closed) return; + + // 同一次反馈的第二屏：只换正文，避免对话框闪一下丢焦点 + if (dialog != null && dialog.isShowing() && session.feedbackId().equals(feedbackId)) { + if (body != null) body.setText(describe(owner, session)); + applyButtons(session); + return; + } + + dismissInternal(); + feedbackId = session.feedbackId(); + + ScrollView scroll = new ScrollView(owner); + LinearLayout content = new LinearLayout(owner); + content.setOrientation(LinearLayout.VERTICAL); + int padding = ResUtil.dp2px(20); + content.setPadding(padding, padding, padding, padding); + body = new TextView(owner); + body.setTextSize(14f); + body.setLineSpacing(0f, 1.2f); + body.setText(describe(owner, session)); + content.addView(body, new LinearLayout.LayoutParams( + LinearLayout.LayoutParams.MATCH_PARENT, LinearLayout.LayoutParams.WRAP_CONTENT)); + scroll.addView(content); + + MaterialAlertDialogBuilder builder = new MaterialAlertDialogBuilder(owner) + .setTitle(title(owner, session)) + .setView(scroll) + .setNegativeButton(R.string.ad_interval_dismiss, (d, which) -> d.dismiss()); + if (session.hasActionablePlan()) { + builder.setPositiveButton(R.string.ad_interval_save, (d, which) -> onSave(session)); + } + dialog = builder.create(); + dialog.setOnDismissListener(d -> { + dialog = null; + body = null; + }); + dialog.show(); + centerButtons(); + } + + /** + * 归因完成后可能才出现「保存规则」按钮。AlertDialog 不支持事后加按钮， + * 因此重建一次；此时用户尚未与内容交互，重建代价可接受。 + */ + private void applyButtons(AdFeedbackSession session) { + if (dialog == null) return; + boolean hasSave = dialog.getButton(AlertDialog.BUTTON_POSITIVE) != null + && dialog.getButton(AlertDialog.BUTTON_POSITIVE).getVisibility() == View.VISIBLE; + if (session.hasActionablePlan() == hasSave) return; + feedbackId = null; + showInternal(session); + } + + private void onSave(AdFeedbackSession session) { + if (callback == null || session.verdict() == null) return; + AdAttribution plan = session.verdict().preferred(); + if (plan != null) callback.onSave(plan); + } + + private String title(FragmentActivity owner, AdFeedbackSession session) { + return owner.getString(R.string.ad_interval_title, + clock(session.startMs()), clock(session.endMs())); + } + + private String describe(FragmentActivity owner, AdFeedbackSession session) { + StringBuilder text = new StringBuilder(); + text.append(owner.getString(R.string.ad_interval_duration, session.durationMs() / 1000f)); + text.append('\n').append(originText(owner, session)); + text.append('\n').append(owner.getString(session.skipApplied() + ? R.string.ad_interval_skipped : R.string.ad_interval_recorded)); + + if (!session.analysisComplete()) { + text.append("\n\n").append(owner.getString(R.string.ad_interval_analyzing)); + return text.toString(); + } + + AdAttribution preferred = session.verdict().preferred(); + text.append("\n\n").append(planText(owner, preferred)); + if (preferred != null) { + text.append(" ").append(riskText(owner, preferred.risk())); + text.append(String.format(Locale.US, " %d%%", + Math.round(preferred.confidence() * 100))); + } + + appendLines(text, owner.getString(R.string.ad_interval_evidence), + session.preferredEvidence()); + if (session.verdict() != null && !session.verdict().diagnostics().isEmpty()) { + StringBuilder diagnostics = new StringBuilder(); + for (AdAttribution diagnostic : session.verdict().diagnostics()) { + for (String line : diagnostic.evidence()) { + if (diagnostics.length() > 0) diagnostics.append('\n'); + diagnostics.append("· ").append(line); + } + } + if (diagnostics.length() > 0) { + text.append("\n\n").append(owner.getString(R.string.ad_interval_diagnostics)) + .append('\n').append(diagnostics); + } + } + return text.toString(); + } + + private static void appendLines(StringBuilder text, String header, List lines) { + if (lines == null || lines.isEmpty()) return; + text.append("\n\n").append(header); + for (String line : lines) text.append("\n· ").append(line); + } + + private static String originText(FragmentActivity owner, AdFeedbackSession session) { + StartOrigin origin = session.startOrigin(); + if (origin == StartOrigin.USER_MARKED) return owner.getString(R.string.ad_interval_origin_marked); + if (origin == StartOrigin.DISCONTINUITY) return owner.getString(R.string.ad_interval_origin_discontinuity); + if (origin == StartOrigin.CROSS_DOMAIN) return owner.getString(R.string.ad_interval_origin_cross_domain); + if (origin == StartOrigin.AUDIO_CANDIDATE) return owner.getString(R.string.ad_interval_origin_audio); + return owner.getString(R.string.ad_interval_origin_window, + (int) TimeUnit.MILLISECONDS.toSeconds(session.durationMs())); + } + + private static String planText(FragmentActivity owner, AdAttribution plan) { + if (plan == null) return owner.getString(R.string.ad_interval_plan_none); + RemediationKind kind = plan.remediation(); + if (kind == RemediationKind.ENABLE_EXISTING_RULE) return owner.getString(R.string.ad_interval_plan_enable); + if (kind == RemediationKind.HOST_BLACKLIST) return owner.getString(R.string.ad_interval_plan_host); + if (kind == RemediationKind.HLS_STRUCTURED_RULE) return owner.getString(R.string.ad_interval_plan_hls); + if (kind == RemediationKind.URL_REGEX_RULE) return owner.getString(R.string.ad_interval_plan_regex); + return owner.getString(R.string.ad_interval_plan_none); + } + + private static String riskText(FragmentActivity owner, RiskLevel risk) { + if (risk == RiskLevel.LOW) return owner.getString(R.string.ad_interval_risk_low); + if (risk == RiskLevel.HIGH) return owner.getString(R.string.ad_interval_risk_high); + return owner.getString(R.string.ad_interval_risk_medium); + } + + /** mm:ss，超过一小时用 h:mm:ss。 */ + static String clock(long millis) { + long total = Math.max(0L, millis) / 1000L; + long hours = total / 3600L; + long minutes = (total % 3600L) / 60L; + long seconds = total % 60L; + return hours > 0 + ? String.format(Locale.US, "%d:%02d:%02d", hours, minutes, seconds) + : String.format(Locale.US, "%02d:%02d", minutes, seconds); + } + + /** TV 上按钮默认贴右，遥控器操作时居中更容易命中。 */ + private void centerButtons() { + if (dialog == null || dialog.getWindow() == null) return; + View parent = dialog.getWindow().findViewById(android.R.id.button1); + if (parent != null && parent.getParent() instanceof LinearLayout buttons) { + buttons.setGravity(Gravity.CENTER_HORIZONTAL); + } + } + + private void dismissInternal() { + if (dialog != null) { + dialog.setOnDismissListener(null); + dialog.dismiss(); + } + dialog = null; + body = null; + } + + private FragmentActivity usableActivity() { + FragmentActivity owner = activity.get(); + if (owner == null || owner.isFinishing() || owner.isDestroyed()) return null; + return owner; + } + + private void runOnMain(Runnable task) { + if (Looper.myLooper() == Looper.getMainLooper()) task.run(); + else main.post(task); + } +} diff --git a/app/src/main/res/values-zh-rCN/strings.xml b/app/src/main/res/values-zh-rCN/strings.xml index 8f7993a803..35873279aa 100644 --- a/app/src/main/res/values-zh-rCN/strings.xml +++ b/app/src/main/res/values-zh-rCN/strings.xml @@ -2470,6 +2470,31 @@ 白名单（正则） （无） 应用并屏蔽 + 广告区间 %1$s – %2$s + 时长 %1$.1f 秒 + 本次播放已跳过 + 已记录。该区间已经播过，未执行跳过。 + 起点：由你框选 + 起点：根据切片断点推断 + 起点：根据切片跨域切换推断 + 起点：复用音频检测结果 + 起点：按回溯 %1$d 秒估算 + 正在分析… + 新增 HLS 切片规则 + 拉黑广告域名 + 新增 URL 正则规则 + 启用已有规则 + 未能归纳出可复用的规则 + 风险：低 + 风险：中 + 风险：高 + 证据 + 诊断 + 保存规则 + 关闭 + 已标记广告起点，到结束处再按一次 + 已取消标记 + 当前位置无法反馈广告 AI/自定义 %1$d 条 · 接口配置 %2$d 条 · 内置 HLS %3$d 条 内置 HLS · v%1$d · %2$s 已启用 diff --git a/app/src/main/res/values-zh-rTW/strings.xml b/app/src/main/res/values-zh-rTW/strings.xml index 5c8cedf5af..fe9fb7816e 100644 --- a/app/src/main/res/values-zh-rTW/strings.xml +++ b/app/src/main/res/values-zh-rTW/strings.xml @@ -2423,6 +2423,31 @@ 白名單（正則） （無） 應用並屏蔽 + 廣告區間 %1$s – %2$s + 時長 %1$.1f 秒 + 本次播放已跳過 + 已記錄。該區間已經播過，未執行跳過。 + 起點：由你框選 + 起點：根據切片斷點推斷 + 起點：根據切片跨域切換推斷 + 起點：複用音訊檢測結果 + 起點：按回溯 %1$d 秒估算 + 正在分析… + 新增 HLS 切片規則 + 拉黑廣告域名 + 新增 URL 正則規則 + 啟用已有規則 + 未能歸納出可複用的規則 + 風險：低 + 風險：中 + 風險：高 + 證據 + 診斷 + 儲存規則 + 關閉 + 已標記廣告起點，到結束處再按一次 + 已取消標記 + 當前位置無法反饋廣告 AI/自定義 %1$d 條 · 接口配置 %2$d 條 · 內置 HLS %3$d 條 內置 HLS · v%1$d · %2$s 已啓用 diff --git a/app/src/main/res/values/strings.xml b/app/src/main/res/values/strings.xml index 0181271901..0fc0cb8c5a 100644 --- a/app/src/main/res/values/strings.xml +++ b/app/src/main/res/values/strings.xml @@ -2137,6 +2137,31 @@ Allowlist (regex) (none) Apply and block + Ad interval %1$s – %2$s + Length %1$.1fs + Skipped for this playback + Recorded. The interval already played, so nothing was skipped. + start: marked by you + start: inferred from segment discontinuity + start: inferred from a cross-domain segment switch + start: reused from an audio detection + start: estimated by rewinding %1$ds + Analyzing… + Add an HLS segment rule + Block the ad domain + Add a URL pattern rule + Turn on an existing rule + No reusable rule could be derived + risk: low + risk: medium + risk: high + Evidence + Diagnostics + Save rule + Close + Ad start marked. Press again at the end. + Marking cancelled + Cannot report an ad at this position AI/custom %1$d · source config %2$d · built-in HLS %3$d Built-in HLS · v%1$d · %2$s Enabled From 9b2c31f1e0372e121f994845fa2294722560f035 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 01:02:35 +0800 Subject: [PATCH 11/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E5=BD=92?= =?UTF-8?q?=E5=9B=A0=E7=BB=93=E8=AE=BA=E6=90=BA=E5=B8=A6=E6=9C=BA=E5=99=A8?= =?UTF-8?q?=E5=8F=AF=E8=AF=BB=E7=9A=84=E8=A7=84=E5=88=99=E8=BD=BD=E8=8D=B7?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第十一个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 6.2、10 节。 接线 UI 时发现一个设计缺口：AdAttribution 只有给人看的 evidence 字符串， 没有落地规则所需的结构化数据。若从证据文本里正则提取域名，规则会随文案 和语言变化而失效 —— 这是不可接受的耦合。 - RulePayload：域名/正则/排除/规则键/时长范围/信号门限。三个分类器各自 填充自己那部分，HLS 通道额外带 playlistHostSuffixes 把作用域收窄到当前 站点，满足设计文档第 10 节「禁止生成全局生效的规则」。 - AdAttribution.actionable() 现在要求载荷非空，否则「保存规则」会写出空规则。 - 仲裁合并同类结论时载荷跟随被选中的机制，不能拿 A 的数据执行 B 的动作。 - 时长范围仅在时长离群信号成立时给出，避免只靠固定时长删片； minimumSignals 恒不低于 2。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback: 109 tests, 0 failures (RulePayloadTest 10 new); compileMobileArm64_v8aDebug OK Task-Guard: AF4g1-payload --- .../android/tv/ad/feedback/AdAttribution.java | 19 ++- .../tv/ad/feedback/AdAttributionArbiter.java | 4 +- .../feedback/DomainReputationClassifier.java | 3 +- .../ad/feedback/ExistingRuleClassifier.java | 3 +- .../tv/ad/feedback/HlsSegmentClassifier.java | 31 +++- .../android/tv/ad/feedback/RulePayload.java | 84 ++++++++++ .../ad/feedback/AdAttributionArbiterTest.java | 17 +- .../tv/ad/feedback/RulePayloadTest.java | 157 ++++++++++++++++++ 8 files changed, 310 insertions(+), 8 deletions(-) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/RulePayload.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttribution.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttribution.java index 47f6e145ab..8c02b0adab 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttribution.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttribution.java @@ -11,10 +11,12 @@ import java.util.List; * @param risk 采纳该结论的风险等级 * @param evidence 人类可读的证据，直接进 UI * @param remediation 建议的去广机制 + * @param payload 落地规则所需的机器可读数据 */ public record AdAttribution( String channelId, AdCategory category, float confidence, - RiskLevel risk, List evidence, RemediationKind remediation) { + RiskLevel risk, List evidence, RemediationKind remediation, + RulePayload payload) { public AdAttribution { if (channelId == null || channelId.isBlank()) { @@ -25,6 +27,13 @@ public record AdAttribution( risk = risk == null ? RiskLevel.MEDIUM : risk; evidence = evidence == null ? List.of() : List.copyOf(evidence); remediation = remediation == null ? RemediationKind.NONE : remediation; + payload = payload == null ? RulePayload.empty() : payload; + } + + /** 纯诊断结论没有可落地的数据。 */ + public AdAttribution(String channelId, AdCategory category, float confidence, + RiskLevel risk, List evidence, RemediationKind remediation) { + this(channelId, category, confidence, risk, evidence, remediation, RulePayload.empty()); } private static float clamp(float value) { @@ -41,9 +50,13 @@ public record AdAttribution( return confidence * 0.7f + remediation.priorityWeight() * 0.3f; } - /** 是否能产出可落地的规则。 */ + /** + * 是否能产出可落地的规则。除了机制本身要可执行，载荷也必须非空 —— + * 否则「保存规则」会写出一条空规则。 + */ public boolean actionable() { return remediation != RemediationKind.NONE - && remediation != RemediationKind.SESSION_SKIP_ONLY; + && remediation != RemediationKind.SESSION_SKIP_ONLY + && !payload.isEmpty(); } } diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiter.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiter.java index 832ee1ae36..badb414d91 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiter.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiter.java @@ -111,7 +111,9 @@ public final class AdAttributionArbiter { channels.add(attribution.channelId()); } merged.add(new AdAttribution(String.join("+", channels), entry.getKey(), - confidence, risk, List.copyOf(lines), cheapest.remediation())); + confidence, risk, List.copyOf(lines), cheapest.remediation(), + // 载荷必须跟随被选中的机制，否则会拿 A 的数据去执行 B 的动作 + cheapest.payload())); } return merged; } diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java index f9fa403c5b..bfacee1a63 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java @@ -91,7 +91,8 @@ public final class DomainReputationClassifier { evidenceLines.add("起点来源：" + evidence.startOrigin()); return new AdAttribution(CHANNEL_ID, AdCategory.THIRD_PARTY_CDN_SEGMENT, - confidence, RiskLevel.LOW, evidenceLines, RemediationKind.HOST_BLACKLIST); + confidence, RiskLevel.LOW, evidenceLines, RemediationKind.HOST_BLACKLIST, + RulePayload.ofHosts(List.copyOf(foreignHosts))); } /** 区间内不属于 playlist 域名的切片 host，去重保序。 */ diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java index 2bf296d0d3..863dc6514e 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java @@ -74,7 +74,8 @@ public final class ExistingRuleClassifier { lines.add("启用它即可拦掉这段广告，无需新建规则"); return new AdAttribution(CHANNEL_ID, AdCategory.DISCONTINUITY_BLOCK, CONFIDENCE_DISABLED_RULE, RiskLevel.LOW, lines, - RemediationKind.ENABLE_EXISTING_RULE); + RemediationKind.ENABLE_EXISTING_RULE, + RulePayload.ofRuleKey(disabled.key())); } List lines = new ArrayList<>(); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index d17493f823..62c2eda652 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -80,7 +80,36 @@ public final class HlsSegmentClassifier { if (signals.encodableCount() == 0) return null; return new AdAttribution(CHANNEL_ID, categoryOf(signals), confidence, - RiskLevel.MEDIUM, describe(evidence, signals), RemediationKind.HLS_STRUCTURED_RULE); + RiskLevel.MEDIUM, describe(evidence, signals), RemediationKind.HLS_STRUCTURED_RULE, + payloadOf(evidence, signals)); + } + + /** + * 生成可被 {@code HlsManifestCleaner.Rule} 执行的条件。 + * + *

时长范围围绕区间内均值取 ±0.1s，仅在时长离群信号成立时给出 —— + * 否则会写出一条只靠固定时长删片的宽泛规则，这是内置规则策略明确禁止的。 + * minimumSignals 取可编码信号数且不低于 2。 + */ + private static RulePayload payloadOf(AdIntervalEvidence evidence, Signals signals) { + List hosts = evidence.inside().stream() + .map(SegmentFact::host) + .filter(host -> !host.isEmpty()) + .distinct() + .toList(); + double min = Double.NaN; + double max = Double.NaN; + if (signals.durationOutlier()) { + double mean = mean(evidence.inside()); + min = Math.max(0d, mean - 0.1d); + max = mean + 0.1d; + } + // 作用域必须收窄到当前站点的 playlist 域名，否则规则会污染其他站点 + List scope = evidence.playlistHost().isEmpty() + ? List.of() : List.of(evidence.playlistHost()); + return RulePayload.ofHlsRule(scope, hosts, min, max, + signals.discontinuity(), signals.crossDomain(), + Math.max(2, signals.encodableCount())); } /** 逐项检测五种信号。 */ diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RulePayload.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RulePayload.java new file mode 100644 index 0000000000..12d8131fa6 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RulePayload.java @@ -0,0 +1,84 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.List; + +/** + * 归因结论的机器可读载荷：落地规则所需的实际数据。 + * + *

与 {@link AdAttribution#evidence()} 分开是刻意的 —— 证据是给人看的、 + * 会随文案和语言变化，不能用来提取域名或规则键。 + * + * @param playlistHostSuffixes 规则作用域，限定到当前站点的 playlist 域名； + * HLS 规则必须携带，否则会污染其他站点（设计文档第 10 节） + * @param hosts 要拉黑或匹配的切片域名 + * @param regex 广告 URL 正则 + * @param exclude 正片保护正则 + * @param ruleKey 要启用的已有规则的复合状态键 + * @param durationMin 切片时长下界，NaN 表示不限定 + * @param durationMax 切片时长上界，NaN 表示不限定 + * @param requireDiscontinuity 是否要求断点边界 + * @param requireCrossDomain 是否要求跨域 + * @param minimumSignals 生成 HLS 规则时的多信号门限 + */ +public record RulePayload(List playlistHostSuffixes, List hosts, + List regex, List exclude, + String ruleKey, double durationMin, double durationMax, + boolean requireDiscontinuity, boolean requireCrossDomain, + int minimumSignals) { + + public RulePayload { + playlistHostSuffixes = playlistHostSuffixes == null ? List.of() : List.copyOf(playlistHostSuffixes); + hosts = hosts == null ? List.of() : List.copyOf(hosts); + regex = regex == null ? List.of() : List.copyOf(regex); + exclude = exclude == null ? List.of() : List.copyOf(exclude); + ruleKey = ruleKey == null ? "" : ruleKey; + if (minimumSignals < 0) minimumSignals = 0; + } + + public static RulePayload empty() { + return new RulePayload(List.of(), List.of(), List.of(), List.of(), "", + Double.NaN, Double.NaN, false, false, 0); + } + + /** 仅域名黑名单。 */ + public static RulePayload ofHosts(List hosts) { + return new RulePayload(List.of(), hosts, List.of(), List.of(), "", + Double.NaN, Double.NaN, false, false, 0); + } + + /** 仅 URL 正则。 */ + public static RulePayload ofRegex(List regex) { + return new RulePayload(List.of(), List.of(), regex, List.of(), "", + Double.NaN, Double.NaN, false, false, 0); + } + + /** 启用一条已有规则。 */ + public static RulePayload ofRuleKey(String ruleKey) { + return new RulePayload(List.of(), List.of(), List.of(), List.of(), ruleKey, + Double.NaN, Double.NaN, false, false, 0); + } + + /** HLS 结构化条件，不带时长范围。 */ + public static RulePayload ofHls(List playlistHostSuffixes, List hosts, + int minimumSignals) { + return new RulePayload(playlistHostSuffixes, hosts, List.of(), List.of(), "", + Double.NaN, Double.NaN, false, false, minimumSignals); + } + + /** HLS 结构化条件全量。 */ + public static RulePayload ofHlsRule(List playlistHostSuffixes, List hosts, + double durationMin, double durationMax, + boolean requireDiscontinuity, boolean requireCrossDomain, + int minimumSignals) { + return new RulePayload(playlistHostSuffixes, hosts, List.of(), List.of(), "", + durationMin, durationMax, requireDiscontinuity, requireCrossDomain, minimumSignals); + } + + public boolean hasDurationRange() { + return !Double.isNaN(durationMin) && !Double.isNaN(durationMax); + } + + public boolean isEmpty() { + return hosts.isEmpty() && regex.isEmpty() && exclude.isEmpty() && ruleKey.isEmpty(); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiterTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiterTest.java index b09c7c8de6..cfcac83c58 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiterTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdAttributionArbiterTest.java @@ -15,7 +15,22 @@ public class AdAttributionArbiterTest { private static AdAttribution of(String channel, AdCategory category, float confidence, RemediationKind kind) { return new AdAttribution(channel, category, confidence, RiskLevel.MEDIUM, - List.of(channel + " 证据"), kind); + List.of(channel + " 证据"), kind, payloadFor(kind)); + } + + /** + * 可落地的机制必须带非空载荷，否则 {@link AdAttribution#actionable()} 为 false + * 并被仲裁器降级成诊断。诊断类机制本身就没有载荷。 + */ + private static RulePayload payloadFor(RemediationKind kind) { + return switch (kind) { + case ENABLE_EXISTING_RULE -> RulePayload.ofRuleKey("builtin|pkg|rule"); + case HOST_BLACKLIST -> RulePayload.ofHosts(List.of("ad.example.com")); + case HLS_STRUCTURED_RULE -> RulePayload.ofHls( + List.of("v.example.com"), List.of("ad.example.com"), 2); + case URL_REGEX_RULE -> RulePayload.ofRegex(List.of(".*/ad/.*")); + default -> RulePayload.empty(); + }; } @Test diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java new file mode 100644 index 0000000000..adfe99a69b --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java @@ -0,0 +1,157 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.List; + +public class RulePayloadTest { + + @Test + public void emptyPayloadIsNotActionable() { + AdAttribution plan = new AdAttribution("hls", AdCategory.DISCONTINUITY_BLOCK, 0.9f, + RiskLevel.MEDIUM, List.of("证据"), RemediationKind.HLS_STRUCTURED_RULE, + RulePayload.empty()); + + // 机制可执行但没有数据可写，必须视为不可落地 + assertFalse(plan.actionable()); + } + + @Test + public void diagnosticConstructorDefaultsToEmptyPayload() { + AdAttribution plan = new AdAttribution("domain", AdCategory.ALREADY_HANDLED, 0.95f, + RiskLevel.LOW, List.of("已在黑名单"), RemediationKind.NONE); + + assertTrue(plan.payload().isEmpty()); + assertFalse(plan.actionable()); + } + + @Test + public void hostPayloadMakesPlanActionable() { + AdAttribution plan = new AdAttribution("domain", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.6f, + RiskLevel.LOW, List.of("证据"), RemediationKind.HOST_BLACKLIST, + RulePayload.ofHosts(List.of("ad.example.com"))); + + assertTrue(plan.actionable()); + assertEquals(List.of("ad.example.com"), plan.payload().hosts()); + } + + @Test + public void ruleKeyPayloadIsNotEmpty() { + RulePayload payload = RulePayload.ofRuleKey("builtin|pkg|baofeng"); + + assertFalse(payload.isEmpty()); + assertEquals("builtin|pkg|baofeng", payload.ruleKey()); + } + + @Test + public void durationRangePresenceIsExplicit() { + assertFalse(RulePayload.ofHosts(List.of("a.com")).hasDurationRange()); + assertTrue(RulePayload.ofHlsRule(List.of("v.a.com"), List.of("ad.a.com"), + 6.3, 6.5, true, true, 2).hasDurationRange()); + } + + @Test + public void nullFieldsBecomeEmptyCollections() { + RulePayload payload = new RulePayload(null, null, null, null, null, + Double.NaN, Double.NaN, false, false, -5); + + assertTrue(payload.playlistHostSuffixes().isEmpty()); + assertTrue(payload.hosts().isEmpty()); + assertTrue(payload.regex().isEmpty()); + assertTrue(payload.exclude().isEmpty()); + assertEquals("", payload.ruleKey()); + assertEquals(0, payload.minimumSignals()); + assertTrue(payload.isEmpty()); + } + + @Test + public void hlsClassifierScopesRuleToPlaylistHostAndKeepsMinimumTwoSignals() { + List inside = List.of( + new SegmentFact(10, "ad-cdn.other.com", "/seg/10.ts", 6.4, true), + new SegmentFact(11, "ad-cdn.other.com", "/seg/11.ts", 6.4, false)); + List outside = List.of( + new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false), + new SegmentFact(12, "v.example.com", "/seg/12.ts", 8.0, false)); + AdIntervalEvidence evidence = new AdIntervalEvidence( + 80_000, 92_800, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + "v.example.com", "/play/index.m3u8", true, + inside, outside, true, true, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence); + RulePayload payload = plan.payload(); + + // 作用域收窄到本站，避免规则污染其他站点 + assertEquals(List.of("v.example.com"), payload.playlistHostSuffixes()); + assertEquals(List.of("ad-cdn.other.com"), payload.hosts()); + assertTrue(payload.requireDiscontinuity()); + assertTrue(payload.requireCrossDomain()); + // 永不生成 minimumSignals=1 的宽泛规则 + assertTrue(payload.minimumSignals() >= 2); + assertTrue(plan.actionable()); + } + + @Test + public void durationRangeOnlyAppearsWhenOutlierSignalHolds() { + // 区间内外时长一致：不给时长条件，避免只靠固定时长删片 + List inside = List.of( + new SegmentFact(10, "ad-cdn.other.com", "/ads/10.ts", 8.0, true), + new SegmentFact(11, "ad-cdn.other.com", "/ads/11.ts", 8.0, false)); + List outside = List.of( + new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false), + new SegmentFact(12, "v.example.com", "/seg/12.ts", 8.0, false)); + AdIntervalEvidence evidence = new AdIntervalEvidence( + 80_000, 96_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + "v.example.com", "/play/index.m3u8", true, + inside, outside, true, true, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + RulePayload payload = HlsSegmentClassifier.classify(evidence).payload(); + + assertFalse(payload.hasDurationRange()); + } + + @Test + public void existingRuleAttributionCarriesRuleKey() { + ExistingRuleClassifier.RuleState disabled = new ExistingRuleClassifier.RuleState( + "builtin|pkg|baofeng", "baofeng", "暴风片头", false, true, + List.of("ad-cdn.other.com")); + AdIntervalEvidence evidence = new AdIntervalEvidence( + 10_000, 40_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + "v.example.com", "/play/index.m3u8", true, + List.of(new SegmentFact(3, "ad-cdn.other.com", "/seg/3.ts", 6.4, true)), + List.of(new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false)), + false, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + AdAttribution plan = ExistingRuleClassifier.classify(evidence, + new ExistingRuleClassifier.Input(List.of(disabled), false, List.of())); + + assertEquals("builtin|pkg|baofeng", plan.payload().ruleKey()); + assertTrue(plan.actionable()); + } + + @Test + public void mergedAttributionKeepsPayloadOfChosenMechanism() { + AdAttribution hls = new AdAttribution("hls", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.6f, + RiskLevel.MEDIUM, List.of("hls 证据"), RemediationKind.HLS_STRUCTURED_RULE, + RulePayload.ofHls(List.of("v.example.com"), List.of("hls-host.com"), 2)); + AdAttribution domain = new AdAttribution("domain", AdCategory.THIRD_PARTY_CDN_SEGMENT, 0.5f, + RiskLevel.LOW, List.of("domain 证据"), RemediationKind.HOST_BLACKLIST, + RulePayload.ofHosts(List.of("domain-host.com"))); + + AdAttributionArbiter.Verdict verdict = AdAttributionArbiter.arbitrate(List.of(hls, domain)); + AdAttribution preferred = verdict.preferred(); + + // 合并后取成本更低的 HOST_BLACKLIST，载荷必须同步取它的，不能错配 + assertEquals(RemediationKind.HOST_BLACKLIST, preferred.remediation()); + assertEquals(List.of("domain-host.com"), preferred.payload().hosts()); + } +} From f5d80b1de097eeaf6ef5c5a711fb30411821bf28 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 01:34:28 +0800 Subject: [PATCH 12/28] =?UTF-8?q?=E6=96=B0=E5=A2=9E=EF=BC=9A=E8=A7=84?= =?UTF-8?q?=E5=88=99=E8=90=BD=E5=9C=B0=E5=99=A8=E4=B8=8E=E7=94=A8=E6=88=B7?= =?UTF-8?q?=20HLS=20=E8=A7=84=E5=88=99=E5=BA=93?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 第十二个逻辑单元，见 docs/ad-feedback-timerange-design.md 第 6.2 节。 写落地器时查证了一件关键事实：RuleConfig.getAds() 的唯一消费者是 CustomWebView，即域名黑名单只拦 WebView 请求，拦不住播放器直连的切片请求。 所以域名黑名单对区间广告是无效落地 —— 这正是设计文档 7.2 节把「已在黑名单 却仍被看到」标记为 ALREADY_HANDLED 的原因。 - UserHlsRuleStore：新增用户可写的 HLS 规则来源并接入 HlsRuleConfig。 内置规则来自 APK assets、接口规则来自点播/直播配置，两者都不可写， 而区间反馈需要产出真正能删除切片的规则。写入前先 compile 校验， 非法规则直接拒绝，避免在 HlsRuleConfig 里留下永久报错条目。 - HlsAdRule.createUserRule()：字段全私有无 setter，补一个工厂方法。 - DomainReputationClassifier 的落地机制从 HOST_BLACKLIST 改为 HLS_STRUCTURED_RULE，携带 playlistHostSuffixes 收窄作用域； hostSuffixes + requireCrossDomain 两个信号刚好满足 minimumSignals=2。 取不到 playlist 域名时才退回域名黑名单。 - AdRulePlanApplier：三条落地路径（启用已有规则/HLS 规则/用户规则）， 数据全部取自 RulePayload，不解析证据文本。返回 Outcome 三态而非 boolean，让 UI 能区分「跳过」和「失败」。 RulePayloadCompileContractTest 锁定最易断的一环：分类器产出的载荷必须能 组装成 compile() 接受的规则。compile() 要求 minimumSignals 不超过实际信号 数，若不满足会出现「已保存但广告依旧」的静默失败。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback: 114 tests, 0 failures (RulePayloadCompileContractTest 5 new); compileMobileArm64_v8aDebug OK Task-Guard: AF4g2-applier --- .../tv/ad/feedback/AdRulePlanApplier.java | 96 +++++++++++++ .../feedback/DomainReputationClassifier.java | 15 +- .../android/tv/api/config/HlsRuleConfig.java | 1 + .../tv/api/config/UserHlsRuleStore.java | 78 +++++++++++ .../com/fongmi/android/tv/bean/HlsAdRule.java | 29 ++++ .../DomainReputationClassifierTest.java | 11 +- .../RulePayloadCompileContractTest.java | 130 ++++++++++++++++++ 7 files changed, 357 insertions(+), 3 deletions(-) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/AdRulePlanApplier.java create mode 100644 app/src/main/java/com/fongmi/android/tv/api/config/UserHlsRuleStore.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadCompileContractTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdRulePlanApplier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdRulePlanApplier.java new file mode 100644 index 0000000000..94e6df8cf4 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdRulePlanApplier.java @@ -0,0 +1,96 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.fongmi.android.tv.api.config.HlsRuleStateStore; +import com.fongmi.android.tv.api.config.UserAdRuleStore; +import com.fongmi.android.tv.api.config.UserHlsRuleStore; +import com.fongmi.android.tv.bean.HlsAdRule; +import com.fongmi.android.tv.bean.UserAdRule; + +import java.util.List; +import java.util.Locale; +import java.util.UUID; + +/** + * 把仲裁选出的方案写成真实规则。 + * + *

数据全部取自 {@link RulePayload}，不从证据文本里解析 —— 证据是给人看的、 + * 随语言变化。见设计文档第 6.2 节。 + */ +public final class AdRulePlanApplier { + + private AdRulePlanApplier() { + } + + /** 落地结果。 */ + public enum Outcome { + /** 已写入并立即生效。 */ + APPLIED, + /** 方案不可落地或数据为空。 */ + SKIPPED, + /** 规则非法或写入失败。 */ + FAILED + } + + /** + * @param plan 仲裁首选方案 + * @param siteKey 当前站点，用于收窄用户规则作用域 + */ + public static Outcome apply(AdAttribution plan, String siteKey) { + if (plan == null || !plan.actionable()) return Outcome.SKIPPED; + try { + return switch (plan.remediation()) { + case ENABLE_EXISTING_RULE -> enableExisting(plan.payload()); + case HLS_STRUCTURED_RULE -> saveHlsRule(plan, siteKey); + case HOST_BLACKLIST -> saveUserRule(plan, siteKey, true); + case URL_REGEX_RULE -> saveUserRule(plan, siteKey, false); + default -> Outcome.SKIPPED; + }; + } catch (RuntimeException e) { + return Outcome.FAILED; + } + } + + private static Outcome enableExisting(RulePayload payload) { + if (payload.ruleKey().isEmpty()) return Outcome.SKIPPED; + HlsRuleStateStore.setEnabled(payload.ruleKey(), true); + return Outcome.APPLIED; + } + + /** + * 写入一条用户 HLS 规则。这是唯一能真正删除广告切片的机制 —— + * 域名黑名单只在 WebView 层生效，不拦播放器直连的切片请求。 + */ + private static Outcome saveHlsRule(AdAttribution plan, String siteKey) { + RulePayload payload = plan.payload(); + if (payload.hosts().isEmpty() && payload.regex().isEmpty()) return Outcome.SKIPPED; + // 没有作用域的规则会跨站点生效，HlsAdRule.compile() 也会拒绝 + if (payload.playlistHostSuffixes().isEmpty()) return Outcome.SKIPPED; + HlsAdRule rule = HlsAdRule.createUserRule( + UUID.randomUUID().toString(), name(plan, siteKey), + payload.playlistHostSuffixes(), payload.hosts(), payload.regex(), + payload.hasDurationRange() ? payload.durationMin() : null, + payload.hasDurationRange() ? payload.durationMax() : null, + payload.requireDiscontinuity(), payload.requireCrossDomain(), + Math.max(2, payload.minimumSignals())); + return UserHlsRuleStore.add(rule) ? Outcome.APPLIED : Outcome.FAILED; + } + + private static Outcome saveUserRule(AdAttribution plan, String siteKey, boolean hostMode) { + RulePayload payload = plan.payload(); + List values = hostMode ? payload.hosts() : payload.regex(); + if (values.isEmpty()) return Outcome.SKIPPED; + UserAdRule rule = UserAdRule.createManual(name(plan, siteKey)); + rule.setSource(UserAdRule.SOURCE_AI); + rule.setSiteKey(siteKey); + if (hostMode) rule.setHosts(values); + else rule.setRegex(values); + UserAdRuleStore.add(rule); + return Outcome.APPLIED; + } + + private static String name(AdAttribution plan, String siteKey) { + String site = siteKey == null || siteKey.isBlank() ? "unknown" : siteKey; + return String.format(Locale.US, "区间反馈 %s %s", + site, plan.category().name().toLowerCase(Locale.US).replace('_', '-')); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java index bfacee1a63..6db88411b6 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java @@ -90,9 +90,20 @@ public final class DomainReputationClassifier { } evidenceLines.add("起点来源：" + evidence.startOrigin()); + // 落地方式是 HLS 结构化规则而不是域名黑名单：黑名单只在 WebView + // 请求拦截生效（RuleConfig.getAds() 的唯一消费者是 CustomWebView）， + // 拦不住播放器直连的切片请求。跨域按 foreignHosts 的定义天然成立， + // 与 hostSuffixes 合起来正好满足 minimumSignals=2。 + List scope = evidence.playlistHost().isEmpty() + ? List.of() : List.of(evidence.playlistHost()); + RemediationKind remediation = scope.isEmpty() + ? RemediationKind.HOST_BLACKLIST : RemediationKind.HLS_STRUCTURED_RULE; + RulePayload payload = scope.isEmpty() + ? RulePayload.ofHosts(List.copyOf(foreignHosts)) + : RulePayload.ofHlsRule(scope, List.copyOf(foreignHosts), + Double.NaN, Double.NaN, false, true, 2); return new AdAttribution(CHANNEL_ID, AdCategory.THIRD_PARTY_CDN_SEGMENT, - confidence, RiskLevel.LOW, evidenceLines, RemediationKind.HOST_BLACKLIST, - RulePayload.ofHosts(List.copyOf(foreignHosts))); + confidence, RiskLevel.LOW, evidenceLines, remediation, payload); } /** 区间内不属于 playlist 域名的切片 host，去重保序。 */ diff --git a/app/src/main/java/com/fongmi/android/tv/api/config/HlsRuleConfig.java b/app/src/main/java/com/fongmi/android/tv/api/config/HlsRuleConfig.java index 27964a31a0..767754cbf1 100644 --- a/app/src/main/java/com/fongmi/android/tv/api/config/HlsRuleConfig.java +++ b/app/src/main/java/com/fongmi/android/tv/api/config/HlsRuleConfig.java @@ -45,6 +45,7 @@ public final class HlsRuleConfig { compileBuiltin(builtin, overrides, compiled, summaries); compileExternal("vod", VodConfig.get().getConfig().getUrl(), VodConfig.get().getHlsRules(), overrides, compiled, summaries); compileExternal("live", LiveConfig.get().getConfig().getUrl(), LiveConfig.get().getHlsRules(), overrides, compiled, summaries); + compileExternal("user", "local", UserHlsRuleStore.load(), overrides, compiled, summaries); rules = List.copyOf(compiled); entries = List.copyOf(summaries); dirty = false; diff --git a/app/src/main/java/com/fongmi/android/tv/api/config/UserHlsRuleStore.java b/app/src/main/java/com/fongmi/android/tv/api/config/UserHlsRuleStore.java new file mode 100644 index 0000000000..d06a2c63e3 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/api/config/UserHlsRuleStore.java @@ -0,0 +1,78 @@ +package com.fongmi.android.tv.api.config; + +import com.fongmi.android.tv.App; +import com.fongmi.android.tv.bean.HlsAdRule; +import com.github.catvod.utils.Prefers; +import com.google.gson.reflect.TypeToken; + +import java.lang.reflect.Type; +import java.util.ArrayList; +import java.util.List; + +/** + * 用户自建的 HLS 结构化规则。 + * + *

内置规则来自 APK assets、接口规则来自点播/直播配置，两者都不可写； + * 而区间反馈需要产出真正能删除广告切片的规则 —— {@code RuleConfig.getAds()} + * 的唯一消费者是 {@code CustomWebView}，只拦 WebView 请求，不拦播放器直连的 + * 切片请求。因此新增这条用户可写的来源。 + */ +public final class UserHlsRuleStore { + + private static final String PREF_KEY = "user_hls_rules"; + private static final Type LIST_TYPE = new TypeToken>() {}.getType(); + /** 上限防止规则库无限膨胀拖慢每次 manifest 匹配。 */ + private static final int MAX_RULES = 200; + + private UserHlsRuleStore() { + } + + public static synchronized List load() { + try { + List rules = App.gson().fromJson(Prefers.getString(PREF_KEY, "[]"), LIST_TYPE); + return rules == null ? new ArrayList<>() : rules; + } catch (Throwable e) { + return new ArrayList<>(); + } + } + + public static synchronized void save(List rules) { + List value = rules == null ? new ArrayList<>() : new ArrayList<>(rules); + while (value.size() > MAX_RULES) value.remove(0); + Prefers.put(PREF_KEY, App.gson().toJson(value)); + HlsRuleConfig.invalidate(); + } + + /** + * 追加一条规则。写入前先 compile 校验，非法规则直接拒绝 —— + * 否则会在 HlsRuleConfig 里变成一条永久报错的条目。 + * + * @return 是否写入成功 + */ + public static synchronized boolean add(HlsAdRule rule) { + if (rule == null) return false; + try { + rule.compile(); + } catch (RuntimeException e) { + return false; + } + List rules = load(); + if (rules.stream().anyMatch(existing -> existing.getId().equals(rule.getId()))) return false; + rules.add(rule); + save(rules); + return true; + } + + public static synchronized boolean delete(String id) { + if (id == null || id.isBlank()) return false; + List rules = load(); + boolean removed = rules.removeIf(rule -> id.equals(rule.getId())); + if (removed) save(rules); + return removed; + } + + public static synchronized void clear() { + Prefers.put(PREF_KEY, "[]"); + HlsRuleConfig.invalidate(); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/bean/HlsAdRule.java b/app/src/main/java/com/fongmi/android/tv/bean/HlsAdRule.java index c0c55a209a..8bb774fe61 100644 --- a/app/src/main/java/com/fongmi/android/tv/bean/HlsAdRule.java +++ b/app/src/main/java/com/fongmi/android/tv/bean/HlsAdRule.java @@ -47,6 +47,35 @@ public class HlsAdRule { return items == null ? Collections.emptyList() : items; } + /** + * 构造一条用户规则。作用域 {@code playlistHostSuffixes} 必填： + * {@link #compile()} 会拒绝没有作用域的规则，避免规则跨站点生效。 + */ + public static HlsAdRule createUserRule(String id, String name, + List playlistHostSuffixes, + List hostSuffixes, + List segmentUrlRegex, + Double minDuration, Double maxDuration, + boolean requireDiscontinuity, + boolean requireCrossDomain, + int minimumSignals) { + HlsAdRule rule = new HlsAdRule(); + rule.id = id; + rule.name = name; + rule.version = 1; + rule.enabledByDefault = true; + rule.enabled = Boolean.TRUE; + rule.playlistHostSuffixes = playlistHostSuffixes; + rule.hostSuffixes = hostSuffixes; + rule.segmentUrlRegex = segmentUrlRegex; + rule.minDuration = minDuration; + rule.maxDuration = maxDuration; + rule.requireDiscontinuity = requireDiscontinuity; + rule.requireCrossDomain = requireCrossDomain; + rule.minimumSignals = minimumSignals; + return rule; + } + public String getId() { return id == null ? "" : id; } diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java index f4000a7759..faf4f2e45f 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java @@ -53,10 +53,19 @@ public class DomainReputationClassifierTest { assertNotNull(attribution); assertEquals(AdCategory.THIRD_PARTY_CDN_SEGMENT, attribution.category()); - assertEquals(RemediationKind.HOST_BLACKLIST, attribution.remediation()); + // 域名黑名单只在 WebView 层生效，拦不住播放器直连的切片； + // 要真正删掉这些切片必须落成 HLS 结构化规则 + assertEquals(RemediationKind.HLS_STRUCTURED_RULE, attribution.remediation()); assertEquals(RiskLevel.LOW, attribution.risk()); assertEquals(DomainReputationClassifier.CONFIDENCE_UNKNOWN_HOST, attribution.confidence(), 0.0001f); + // 规则必须限定到本站 playlist 域名，否则会污染其他站点 + assertEquals(List.of(PLAYLIST_HOST), attribution.payload().playlistHostSuffixes()); + assertEquals(List.of("ad-cdn.other.com"), attribution.payload().hosts()); + // hostSuffixes + requireCrossDomain 两个信号，满足 minimumSignals >= 2 + assertTrue(attribution.payload().requireCrossDomain()); + assertEquals(2, attribution.payload().minimumSignals()); + assertTrue(attribution.actionable()); } @Test diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadCompileContractTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadCompileContractTest.java new file mode 100644 index 0000000000..fd4093c95f --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadCompileContractTest.java @@ -0,0 +1,130 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertNotNull; +import static org.junit.Assert.assertTrue; +import static org.junit.Assert.fail; + +import com.fongmi.android.tv.bean.HlsAdRule; + +import java.util.ArrayList; +import java.util.List; + +import org.junit.Test; + +/** + * 契约测试：分类器产出的 {@link RulePayload} 必须能组装成一条 + * {@code HlsAdRule.compile()} 接受的规则。 + * + *

这是整条链路最容易断的地方 —— 规则写进 UserHlsRuleStore 却编译失败时， + * 用户看到「已保存」但广告依旧存在，且只会在 HlsRuleConfig 里留下一条报错条目。 + * compile() 的硬约束是 minimumSignals 不得超过实际信号数。 + */ +public class RulePayloadCompileContractTest { + + /** 把载荷按 AdRulePlanApplier 的方式组装成规则并编译。 */ + private static void assertCompiles(String label, RulePayload payload) { + assertTrue(label + ": 载荷必须带作用域", !payload.playlistHostSuffixes().isEmpty()); + HlsAdRule rule = HlsAdRule.createUserRule( + "test-id", "test", payload.playlistHostSuffixes(), + payload.hosts(), payload.regex(), + payload.hasDurationRange() ? payload.durationMin() : null, + payload.hasDurationRange() ? payload.durationMax() : null, + payload.requireDiscontinuity(), payload.requireCrossDomain(), + Math.max(2, payload.minimumSignals())); + try { + assertNotNull(label, rule.compile()); + } catch (RuntimeException e) { + fail(label + ": 规则无法编译 -> " + e.getMessage()); + } + } + + private static AdIntervalEvidence evidence(List inside, List outside, + boolean bounded, boolean crossDomain) { + return new AdIntervalEvidence( + 80_000, 100_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + "v.example.com", "/play/index.m3u8", true, + inside, outside, bounded, crossDomain, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + private static List segments(int from, int count, String host, + String pathPrefix, double duration, + boolean discontinuityOnFirst) { + List facts = new ArrayList<>(); + for (int i = 0; i < count; i++) { + facts.add(new SegmentFact(from + i, host, pathPrefix + (from + i) + ".ts", + duration, discontinuityOnFirst && i == 0)); + } + return facts; + } + + @Test + public void crossDomainWithDiscontinuityAndDurationOutlierCompiles() { + AdIntervalEvidence evidence = evidence( + segments(10, 3, "ad-cdn.other.com", "/seg/", 6.4, true), + segments(0, 10, "v.example.com", "/seg/", 8.0, false), + true, true); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence); + + assertNotNull("三信号命中应产出方案", plan); + assertCompiles("跨域+断点+时长离群", plan.payload()); + } + + @Test + public void pathHintOnlyIntervalStillCompiles() { + // 同域、时长一致，只有路径特征命中：可编码信号数为 1， + // 而落地时 minimumSignals 被抬到 2 —— compile() 会因此拒绝。 + AdIntervalEvidence evidence = evidence( + segments(10, 3, "v.example.com", "/ads/", 8.0, false), + segments(0, 10, "v.example.com", "/seg/", 8.0, false), + false, false); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence); + + // 置信度 0.15 低于门槛，通道应弃权而不是产出无法编译的规则 + assertTrue("单一弱信号必须弃权", plan == null); + } + + @Test + public void discontinuityPlusPathHintCompiles() { + AdIntervalEvidence evidence = evidence( + segments(10, 3, "v.example.com", "/ads/", 8.0, true), + segments(0, 10, "v.example.com", "/seg/", 8.0, false), + true, false); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence); + + assertNotNull("断点+路径特征应产出方案", plan); + assertCompiles("断点+路径特征", plan.payload()); + } + + @Test + public void domainChannelPayloadCompiles() { + AdIntervalEvidence evidence = evidence( + segments(10, 3, "ad-cdn.other.com", "/seg/", 6.4, false), + segments(0, 10, "v.example.com", "/seg/", 8.0, false), + false, true); + + AdAttribution plan = DomainReputationClassifier.classify(evidence, + new DomainReputationClassifier.Input( + List.of(), List.of("v.example.com"), List.of(), "")); + + assertNotNull("非本站域名应产出方案", plan); + assertCompiles("域名通道", plan.payload()); + } + + @Test + public void durationOutlierWithCrossDomainCompiles() { + AdIntervalEvidence evidence = evidence( + segments(10, 4, "ad-cdn.other.com", "/seg/", 6.4, false), + segments(0, 10, "v.example.com", "/seg/", 8.0, false), + false, true); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence); + + assertNotNull("跨域+时长离群应产出方案", plan); + assertCompiles("跨域+时长离群", plan.payload()); + } +} From 1bfa7e0ffc532c3bf29c18e1a599deb8e318949a Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 01:55:21 +0800 Subject: [PATCH 13/28] =?UTF-8?q?=E6=8E=A5=E7=BA=BF=EF=BC=9A=E4=B8=89?= =?UTF-8?q?=E7=AB=AF=E3=80=8C=E6=9C=89=E5=B9=BF=E5=91=8A=E3=80=8D=E6=8C=89?= =?UTF-8?q?=E9=92=AE=E6=94=B9=E7=94=A8=E5=8C=BA=E9=97=B4=E5=8F=8D=E9=A6=88?= =?UTF-8?q?=E4=B8=8E=E5=A4=9A=E9=80=9A=E9=81=93=E5=BD=92=E5=9B=A0?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 1 收尾，见 docs/ad-feedback-timerange-design.md 第 4、5.2 节。 这是本系列唯一改变用户可见行为的一步。 三端（leanback / mobile / TmdbDetail）各自那份近乎逐字重复的 submitAdFeedback / buildAdDetectionRequest / enrichRequestWithM3u8Evidence / onAdDetectionResult 全部删除，改为共用 AdFeedbackController + AdFeedbackHostAdapter，每端只留 Playback/Metadata/Ui 三个内部类适配。 行为变化： - 按钮可见性放宽：去掉 isAiConfigReady() && isAiAdDetection() 硬门槛， 改为 isAdblock() && 非直播 && 时长已知。不配 AI 也能用，因为本地归因 通道（切片结构/域名/既有规则）不依赖 AI。 - 短按：以当前位置为终点，按 DISCONTINUITY → 跨域切换 → 音频候选 → 固定窗口逐级回溯推断起点。 - 长按：进入标记模式打起点，再次长按取消，之后短按确定终点。 - 提交后立即跳过该区间并显示第一屏，归因完成后原地替换为第二屏。 - 确认保存时经 AdRulePlanApplier 落成真实规则，三态结果分别提示。 - TmdbDetail 换源换集时作废标记模式并清掉上一集的切片证据缓存。 同步更新 TmdbDetailActivityLayoutTest 的源码断言：原断言锁定了 AI 硬门槛 与 AiAdDetectionService 调用，现改为锁定新契约，并加反向断言确保重复的 per-activity 请求构造器不会回归。 Verification: compileMobileArm64_v8aDebug + compileLeanbackArm64_v8aDebug OK; testMobileArm64_v8aDebugUnitTest ui.activity 348 tests 0 failures; full suite 4002 tests with 5 pre-existing failures (MpvPreloadController/FfmpegVc1Support/MultiThreadProxyPlayerUiSource) verified present on base commit via stash Task-Guard: AF4h-wiring --- .../android/tv/ui/activity/VideoActivity.java | 217 +++++++++++------ .../tv/ui/activity/TmdbDetailActivity.java | 210 +++++++++++------ app/src/main/res/values-zh-rCN/strings.xml | 3 + app/src/main/res/values-zh-rTW/strings.xml | 3 + app/src/main/res/values/strings.xml | 3 + .../android/tv/ui/activity/VideoActivity.java | 219 +++++++++++------- .../TmdbDetailActivityLayoutTest.java | 25 +- 7 files changed, 452 insertions(+), 228 deletions(-) diff --git a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index d70b7a7a59..4fe1295f3a 100644 --- a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -46,8 +46,11 @@ import com.fongmi.android.tv.api.DanmakuApi; import com.fongmi.android.tv.api.SiteApi; import com.fongmi.android.tv.api.config.AdBlockStatsStore; import com.fongmi.android.tv.api.config.VodConfig; -import com.fongmi.android.tv.bean.AdDetectionRequest; -import com.fongmi.android.tv.bean.AdDetectionResult; +import com.fongmi.android.tv.ad.feedback.AdFeedbackController; +import com.fongmi.android.tv.ad.feedback.AdFeedbackHostAdapter; +import com.fongmi.android.tv.ad.feedback.AdFeedbackSession; +import com.fongmi.android.tv.ad.feedback.AdRulePlanApplier; +import com.fongmi.android.tv.ui.dialog.AdFeedbackDialog; import com.fongmi.android.tv.bean.AiConfig; import com.fongmi.android.tv.bean.Danmaku; import com.fongmi.android.tv.bean.Episode; @@ -80,7 +83,6 @@ import com.fongmi.android.tv.player.PlayerHelper; import com.fongmi.android.tv.player.PlayerManager; import com.fongmi.android.tv.player.lut.LutPreset; import com.fongmi.android.tv.player.lut.LutStore; -import com.fongmi.android.tv.service.AiAdDetectionService; import com.fongmi.android.tv.service.PlaybackService; import com.fongmi.android.tv.service.IntroSkipService; import com.fongmi.android.tv.service.OmdbService; @@ -391,7 +393,11 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.recommendationExecutor()); private final Task.Scope mTmdbRatingTasks = new Task.Scope(Task.recommendationExecutor()); private String mTmdbRatingContextKey = ""; - private int mAdFeedbackGeneration; + private AdFeedbackController mAdFeedback; + private AdFeedbackHostAdapter mAdFeedbackHost; + private AdFeedbackDialog mAdFeedbackDialog; + /** 标记模式下已打的起点，-1 表示未进入标记模式。 */ + private long mAdMarkStartMs = -1; // TMDB 模式相关字段 private com.fongmi.android.tv.ui.helper.TmdbUIAdapter mTmdbUIAdapter; @@ -1410,6 +1416,7 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; mBinding.control.action.danmaku.setOnClickListener(guarded(this::onDanmaku)); mBinding.control.action.danmaku.setOnLongClickListener(view -> onDanmakuToggle()); mBinding.control.action.adFeedback.setOnClickListener(view -> onAdFeedback()); + mBinding.control.action.adFeedback.setOnLongClickListener(view -> onAdFeedbackLongPress()); mBinding.control.action.opening.setOnClickListener(guarded(this::onOpening)); if (mBinding.control.action.immersiveAudio != null) mBinding.control.action.immersiveAudio.setOnClickListener(view -> toggleImmersiveAudioMode()); if (mBinding.control.action.cast != null) mBinding.control.action.cast.setOnClickListener(view -> onCast()); @@ -3985,7 +3992,7 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; return; } if (!isAdFeedbackEnabled()) { - Notify.show(R.string.ad_feedback_ai_disabled); + Notify.show(R.string.ad_interval_invalid); return; } hideControl(); @@ -3993,15 +4000,10 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; } private boolean isAdFeedbackEnabled() { - // 功能开关 + 仅支持解析的格式(HLS/m3u8)才可反馈,因为去广分析依赖切片列表 - return Setting.isAiConfigReady() && Setting.isAdblock() && Setting.isAiAdDetection() && isAdFeedbackSupportedFormat(); - } - - private boolean isAdFeedbackSupportedFormat() { - if (player() == null) return false; - String url = player().getUrl(); - if (TextUtils.isEmpty(url)) return false; - return com.fongmi.android.tv.player.exo.MediaSourceFactory.isHlsUrl(url); + // AI 不再是硬门槛：本地归因通道（切片结构/域名/既有规则）不依赖 AI 配置。 + // 只要求可定位的时间轴 —— 直播和时长未知的内容无法框选区间。 + if (player() == null || TextUtils.isEmpty(player().getUrl())) return false; + return Setting.isAdblock() && !player().isLive() && player().getDuration() > 0; } private void setAdFeedbackVisible() { @@ -4009,76 +4011,139 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; applyActionButtonVisibility(); } + private AdFeedbackController adFeedback() { + if (mAdFeedback == null) { + mAdFeedbackHost = new AdFeedbackHostAdapter( + new AdFeedbackPlayback(), new AdFeedbackMetadata(), new AdFeedbackUi()); + mAdFeedback = new AdFeedbackController(mAdFeedbackHost); + } + return mAdFeedback; + } + + /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitAdFeedback() { - AdDetectionRequest request = buildAdDetectionRequest(); - if (request == null) { - Notify.show(R.string.ad_feedback_no_url); - return; + AdFeedbackController controller = adFeedback(); + AdBlockStatsStore.recordFeedback(getKey()); + AdFeedbackSession session; + if (mAdMarkStartMs >= 0) { + long start = mAdMarkStartMs; + mAdMarkStartMs = -1; + session = controller.onMarkedInterval(start, player().getPosition()); + } else { + session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } - // 记录 AI 反馈统计 - AdBlockStatsStore.recordFeedback(request.getSiteKey()); - - Notify.show(R.string.ad_feedback_analyzing); - int generation = ++mAdFeedbackGeneration; - AiConfig config = AiConfig.objectFrom(Setting.getAiConfig()); - Task.execute(() -> { - enrichRequestWithM3u8Evidence(request); - AdDetectionResult result = new AiAdDetectionService(config).analyze(request); - runOnUiThread(() -> { - if (isFinishing() || isDestroyed() || generation != mAdFeedbackGeneration) return; - onAdDetectionResult(request, result); - }); - }); + if (session == null) Notify.show(R.string.ad_interval_invalid); } - private AdDetectionRequest buildAdDetectionRequest() { - if (player() == null || TextUtils.isEmpty(player().getUrl())) return null; - String url = player().getUrl(); - Uri uri = Uri.parse(url); - AdDetectionRequest request = new AdDetectionRequest(); + /** 长按：打起点进入标记模式，再次长按取消。 */ + private boolean onAdFeedbackLongPress() { + if (!isAdFeedbackEnabled()) return false; + if (mAdMarkStartMs >= 0) { + mAdMarkStartMs = -1; + Notify.show(R.string.ad_interval_mark_cancelled); + } else { + mAdMarkStartMs = player().getPosition(); + Notify.show(R.string.ad_interval_mark_start); + } + return true; + } + + private void showAdFeedbackSession(AdFeedbackSession session) { + if (isFinishing() || isDestroyed()) return; + if (mAdFeedbackDialog == null) mAdFeedbackDialog = new AdFeedbackDialog(this, this::applyAdRulePlan); + mAdFeedbackDialog.show(session); + } + + private void applyAdRulePlan(com.fongmi.android.tv.ad.feedback.AdAttribution plan) { Site site = getSite(); - request.setSiteKey(site == null ? getKey() : site.getKey()); - request.setSiteName(site == null ? "" : site.getName()); - request.setVodName(mHistory == null ? getName() : mHistory.getVodName()); - request.setFlagName(getFlag() == null ? "" : getFlag().getFlag()); - request.setEpisodeName(getEpisode() == null ? "" : getEpisode().getName()); - request.setUrlHost(uri.getHost()); - request.setUrlPath(uri.getPath()); - return request; - } - - private void enrichRequestWithM3u8Evidence(AdDetectionRequest request) { - if (player() == null || TextUtils.isEmpty(player().getUrl())) return; - String url = player().getUrl(); - if (!url.contains(".m3u8")) return; - try { - java.util.Map headers = player().getHeaders(); - com.fongmi.android.tv.bean.M3u8Evidence evidence = com.fongmi.android.tv.utils.M3u8Parser.parse(url, headers); - request.setEvidence(evidence); - } catch (Exception e) { - // Ignore parsing failures - } - } - - private void onAdDetectionResult(AdDetectionRequest request, AdDetectionResult result) { - // 记录 AI 分析结果统计 - AdBlockStatsStore.recordAiAnalysis(result != null && !result.isError()); - - if (result == null || result.isError()) { - Notify.show(result == null ? getString(R.string.ad_feedback_failed) : result.getErrorMessage()); - return; - } - if (result.isEmpty()) { - Notify.show(R.string.ad_feedback_no_ad); - return; - } - com.fongmi.android.tv.ui.dialog.AdRulePreviewDialog.create(result).show(this, confirmedResult -> { - UserAdRule rule = UserAdRule.fromAiResult(confirmedResult, request.getSiteKey()); - com.fongmi.android.tv.api.config.UserAdRuleStore.add(rule); - Notify.show(R.string.ad_feedback_saved); + AdRulePlanApplier.Outcome outcome = AdRulePlanApplier.apply( + plan, site == null ? getKey() : site.getKey()); + Notify.show(switch (outcome) { + case APPLIED -> R.string.ad_interval_rule_saved; + case SKIPPED -> R.string.ad_interval_rule_skipped; + case FAILED -> R.string.ad_interval_rule_failed; }); } + private final class AdFeedbackPlayback implements AdFeedbackHostAdapter.Playback { + @Override + public long positionMs() { + return player() == null ? 0 : player().getPosition(); + } + + @Override + public long durationMs() { + return player() == null ? 0 : player().getDuration(); + } + + @Override + public String playUrl() { + return player() == null ? "" : player().getUrl(); + } + + @Override + public java.util.Map headers() { + return player() == null ? java.util.Map.of() : player().getHeaders(); + } + + @Override + public boolean hls() { + return player() != null + && com.fongmi.android.tv.player.exo.MediaSourceFactory.isHlsUrl(player().getUrl()); + } + + @Override + public boolean skipInterval(long startMs, long endMs, String feedbackId) { + return player() != null && player().skipUserAdInterval(startMs, endMs, feedbackId); + } + } + + private final class AdFeedbackMetadata implements AdFeedbackHostAdapter.Metadata { + @Override + public String siteKey() { + Site site = getSite(); + return site == null ? getKey() : site.getKey(); + } + + @Override + public String siteName() { + Site site = getSite(); + return site == null ? "" : site.getName(); + } + + @Override + public String vodName() { + return mHistory == null ? getName() : mHistory.getVodName(); + } + + @Override + public String flagName() { + return getFlag() == null ? "" : getFlag().getFlag(); + } + + @Override + public String episodeName() { + return getEpisode() == null ? "" : getEpisode().getName(); + } + } + + private final class AdFeedbackUi implements AdFeedbackHostAdapter.Ui { + @Override + public void runBackground(Runnable task) { + Task.execute(task); + } + + @Override + public void runOnUi(Runnable task) { + runOnUiThread(task); + } + + @Override + public void showSession(AdFeedbackSession session) { + showAdFeedbackSession(session); + } + } + private boolean onDanmakuToggle() { DanmakuSetting.putShow(!DanmakuSetting.isShow()); applyDanmakuState(); diff --git a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java index 59a94b70a0..65c1be4d59 100644 --- a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java +++ b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java @@ -67,8 +67,11 @@ import com.fongmi.android.tv.api.SiteApi; import com.fongmi.android.tv.api.config.AdBlockStatsStore; import com.fongmi.android.tv.api.config.UserAdRuleStore; import com.fongmi.android.tv.api.config.VodConfig; -import com.fongmi.android.tv.bean.AdDetectionRequest; -import com.fongmi.android.tv.bean.AdDetectionResult; +import com.fongmi.android.tv.ad.feedback.AdFeedbackController; +import com.fongmi.android.tv.ad.feedback.AdFeedbackHostAdapter; +import com.fongmi.android.tv.ad.feedback.AdFeedbackSession; +import com.fongmi.android.tv.ad.feedback.AdRulePlanApplier; +import com.fongmi.android.tv.ui.dialog.AdFeedbackDialog; import com.fongmi.android.tv.bean.AiConfig; import com.fongmi.android.tv.bean.Danmaku; import com.fongmi.android.tv.bean.Episode; @@ -110,7 +113,6 @@ import com.fongmi.android.tv.player.IntroSkipPlayback; import com.fongmi.android.tv.player.PlayerManager; import com.fongmi.android.tv.player.PlayerHelper; import com.fongmi.android.tv.player.exo.MediaSourceFactory; -import com.fongmi.android.tv.service.AiAdDetectionService; import com.fongmi.android.tv.service.AiEpisodeSeasonService; import com.fongmi.android.tv.service.AiRecommendationService; import com.fongmi.android.tv.service.PersonalRecommendationService; @@ -408,7 +410,11 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. /** 本次取详情的起始时刻，用来认出「猫源开内嵌页」是不是自己这次导航触发的。volatile：加载在后台线程发起，事件在主线程读。 */ private volatile long detailLoadStart; private int inlinePlaybackGeneration; - private int mAdFeedbackGeneration; + private AdFeedbackController mAdFeedback; + private AdFeedbackHostAdapter mAdFeedbackHost; + private AdFeedbackDialog mAdFeedbackDialog; + /** 标记模式下已打的起点，-1 表示未进入标记模式。 */ + private long mAdMarkStartMs = -1; private int tmdbDialogGeneration; private AiEpisodeSeasonService aiSeasonService; private AlertDialog aiSeasonLoadingDialog; @@ -1195,6 +1201,7 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. setupInlineControlFocus(); setupInlineFocusNavigation(); binding.playerAdFeedback.setOnClickListener(guarded(this::onInlineAdFeedback)); + binding.playerAdFeedback.setOnLongClickListener(view -> onInlineAdFeedbackLongPress()); binding.playerMultiThreadProxy.setOnClickListener(guarded(this::showInlineMultiThreadProxy)); binding.playerSearch.setOnClickListener(view -> openInlineSourceSearch()); binding.playerSearch.setOnLongClickListener(view -> openGlobalSourceSearch()); @@ -1253,6 +1260,7 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. detailActionView(R.id.ending, View.class).setOnLongClickListener(view -> resetInlineEnding()); detailActionView(R.id.danmaku, View.class).setOnClickListener(guarded(this::showInlineDanmaku)); detailActionView(R.id.adFeedback, View.class).setOnClickListener(guarded(this::onInlineAdFeedback)); + detailActionView(R.id.adFeedback, View.class).setOnLongClickListener(view -> onInlineAdFeedbackLongPress()); detailActionView(R.id.chapter, View.class).setOnClickListener(guarded(this::showInlineTitle)); detailActionView(R.id.episodes, View.class).setOnClickListener(guarded(this::showInlineEpisodes)); setupMobileInlineParse(); @@ -6886,7 +6894,9 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. } private void stopInlinePlayerForReload() { - mAdFeedbackGeneration++; + // 换源换集：作废标记模式并清掉上一集的切片证据缓存 + mAdMarkStartMs = -1; + if (mAdFeedbackHost != null) mAdFeedbackHost.invalidateEvidence(); subtitlePlaybackSession.stop(this); inlineStartPosition = C.TIME_UNSET; inlineStartPositionApplied = false; @@ -8289,7 +8299,7 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. return; } if (!isInlineAdFeedbackEnabled()) { - Notify.show(R.string.ad_feedback_ai_disabled); + Notify.show(R.string.ad_interval_invalid); return; } hideInlineControls(); @@ -8297,79 +8307,145 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. } private boolean isInlineAdFeedbackEnabled() { - return Setting.isAiConfigReady() && Setting.isAdblock() && Setting.isAiAdDetection() - && isInlineAdFeedbackSupportedFormat(); + // AI 不再是硬门槛：本地归因通道（切片结构/域名/既有规则）不依赖 AI 配置。 + // 只要求可定位的时间轴 —— 直播和时长未知的内容无法框选区间。 + if (player() == null || TextUtils.isEmpty(player().getUrl())) return false; + return Setting.isAdblock() && !player().isLive() && player().getDuration() > 0; } - private boolean isInlineAdFeedbackSupportedFormat() { - return player() != null && !TextUtils.isEmpty(player().getUrl()) - && MediaSourceFactory.isHlsUrl(player().getUrl()); + private AdFeedbackController adFeedback() { + if (mAdFeedback == null) { + mAdFeedbackHost = new AdFeedbackHostAdapter( + new AdFeedbackPlayback(), new AdFeedbackMetadata(), new AdFeedbackUi()); + mAdFeedback = new AdFeedbackController(mAdFeedbackHost); + } + return mAdFeedback; } + /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitInlineAdFeedback() { - AdDetectionRequest request = buildInlineAdDetectionRequest(); - if (request == null) { - Notify.show(R.string.ad_feedback_no_url); - return; + AdFeedbackController controller = adFeedback(); + AdBlockStatsStore.recordFeedback(getKeyText()); + AdFeedbackSession session; + if (mAdMarkStartMs >= 0) { + long start = mAdMarkStartMs; + mAdMarkStartMs = -1; + session = controller.onMarkedInterval(start, player().getPosition()); + } else { + session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } - AdBlockStatsStore.recordFeedback(request.getSiteKey()); - Notify.show(R.string.ad_feedback_analyzing); - int generation = ++mAdFeedbackGeneration; - AiConfig config = AiConfig.objectFrom(Setting.getAiConfig()); - detailTasks.submit(Task.recommendationExecutor(), () -> { - enrichInlineAdDetectionRequest(request); - AdDetectionResult result = new AiAdDetectionService(config).analyze(request); - runOnAliveUi(() -> { - if (generation != mAdFeedbackGeneration) return; - onInlineAdDetectionResult(request, result); - }); - }); + if (session == null) Notify.show(R.string.ad_interval_invalid); } - private AdDetectionRequest buildInlineAdDetectionRequest() { - if (player() == null || TextUtils.isEmpty(player().getUrl())) return null; - Uri uri = Uri.parse(player().getUrl()); - AdDetectionRequest request = new AdDetectionRequest(); + /** 长按：打起点进入标记模式，再次长按取消。 */ + private boolean onInlineAdFeedbackLongPress() { + if (!isInlineAdFeedbackEnabled()) return false; + if (mAdMarkStartMs >= 0) { + mAdMarkStartMs = -1; + Notify.show(R.string.ad_interval_mark_cancelled); + } else { + mAdMarkStartMs = player().getPosition(); + Notify.show(R.string.ad_interval_mark_start); + } + return true; + } + + private void showAdFeedbackSession(AdFeedbackSession session) { + if (isFinishing() || isDestroyed()) return; + if (mAdFeedbackDialog == null) mAdFeedbackDialog = new AdFeedbackDialog(this, this::applyAdRulePlan); + mAdFeedbackDialog.show(session); + } + + private void applyAdRulePlan(com.fongmi.android.tv.ad.feedback.AdAttribution plan) { Site site = getCurrentSite(); - History currentHistory = getHistory(); - request.setSiteKey(site == null ? getKeyText() : site.getKey()); - request.setSiteName(site == null ? "" : site.getName()); - request.setVodName(currentHistory == null ? getNameText() : currentHistory.getVodName()); - request.setFlagName(selectedFlag == null ? "" : selectedFlag.getFlag()); - request.setEpisodeName(selectedEpisode == null ? "" : selectedEpisode.getName()); - request.setUrlHost(uri.getHost()); - request.setUrlPath(uri.getPath()); - return request; - } - - private void enrichInlineAdDetectionRequest(AdDetectionRequest request) { - if (player() == null || TextUtils.isEmpty(player().getUrl())) return; - String url = player().getUrl(); - if (!url.contains(".m3u8")) return; - try { - request.setEvidence(com.fongmi.android.tv.utils.M3u8Parser.parse(url, player().getHeaders())); - } catch (Exception ignored) { - // Ignore parsing failures. - } - } - - private void onInlineAdDetectionResult(AdDetectionRequest request, AdDetectionResult result) { - AdBlockStatsStore.recordAiAnalysis(result != null && !result.isError()); - if (result == null || result.isError()) { - Notify.show(result == null ? getString(R.string.ad_feedback_failed) : result.getErrorMessage()); - return; - } - if (result.isEmpty()) { - Notify.show(R.string.ad_feedback_no_ad); - return; - } - AdRulePreviewDialog.create(result).show(this, confirmedResult -> { - UserAdRule rule = UserAdRule.fromAiResult(confirmedResult, request.getSiteKey()); - UserAdRuleStore.add(rule); - Notify.show(R.string.ad_feedback_saved); + AdRulePlanApplier.Outcome outcome = AdRulePlanApplier.apply( + plan, site == null ? getKeyText() : site.getKey()); + Notify.show(switch (outcome) { + case APPLIED -> R.string.ad_interval_rule_saved; + case SKIPPED -> R.string.ad_interval_rule_skipped; + case FAILED -> R.string.ad_interval_rule_failed; }); } + private final class AdFeedbackPlayback implements AdFeedbackHostAdapter.Playback { + @Override + public long positionMs() { + return player() == null ? 0 : player().getPosition(); + } + + @Override + public long durationMs() { + return player() == null ? 0 : player().getDuration(); + } + + @Override + public String playUrl() { + return player() == null ? "" : player().getUrl(); + } + + @Override + public java.util.Map headers() { + return player() == null ? java.util.Map.of() : player().getHeaders(); + } + + @Override + public boolean hls() { + return player() != null && MediaSourceFactory.isHlsUrl(player().getUrl()); + } + + @Override + public boolean skipInterval(long startMs, long endMs, String feedbackId) { + return player() != null && player().skipUserAdInterval(startMs, endMs, feedbackId); + } + } + + private final class AdFeedbackMetadata implements AdFeedbackHostAdapter.Metadata { + @Override + public String siteKey() { + Site site = getCurrentSite(); + return site == null ? getKeyText() : site.getKey(); + } + + @Override + public String siteName() { + Site site = getCurrentSite(); + return site == null ? "" : site.getName(); + } + + @Override + public String vodName() { + History currentHistory = getHistory(); + return currentHistory == null ? getNameText() : currentHistory.getVodName(); + } + + @Override + public String flagName() { + return selectedFlag == null ? "" : selectedFlag.getFlag(); + } + + @Override + public String episodeName() { + return selectedEpisode == null ? "" : selectedEpisode.getName(); + } + } + + private final class AdFeedbackUi implements AdFeedbackHostAdapter.Ui { + @Override + public void runBackground(Runnable task) { + detailTasks.submit(Task.recommendationExecutor(), task); + } + + @Override + public void runOnUi(Runnable task) { + runOnAliveUi(task); + } + + @Override + public void showSession(AdFeedbackSession session) { + showAdFeedbackSession(session); + } + } + private void showInlineDanmaku() { if (service() == null || player().isEmpty()) return; DanmakuDialog.create().player(player()).identity(getKeyText(), getIdText(), playbackHistoryName(), selectedEpisode == null ? "" : historyEpisodeTitle(selectedEpisode)).show(this); diff --git a/app/src/main/res/values-zh-rCN/strings.xml b/app/src/main/res/values-zh-rCN/strings.xml index 35873279aa..cf7b18a824 100644 --- a/app/src/main/res/values-zh-rCN/strings.xml +++ b/app/src/main/res/values-zh-rCN/strings.xml @@ -2495,6 +2495,9 @@ 已标记广告起点，到结束处再按一次 已取消标记 当前位置无法反馈广告 + 规则已保存并生效 + 该区间没有可保存的规则 + 规则保存失败 AI/自定义 %1$d 条 · 接口配置 %2$d 条 · 内置 HLS %3$d 条 内置 HLS · v%1$d · %2$s 已启用 diff --git a/app/src/main/res/values-zh-rTW/strings.xml b/app/src/main/res/values-zh-rTW/strings.xml index fe9fb7816e..4d62df1e00 100644 --- a/app/src/main/res/values-zh-rTW/strings.xml +++ b/app/src/main/res/values-zh-rTW/strings.xml @@ -2448,6 +2448,9 @@ 已標記廣告起點，到結束處再按一次 已取消標記 當前位置無法反饋廣告 + 規則已儲存並生效 + 該區間沒有可儲存的規則 + 規則儲存失敗 AI/自定義 %1$d 條 · 接口配置 %2$d 條 · 內置 HLS %3$d 條 內置 HLS · v%1$d · %2$s 已啓用 diff --git a/app/src/main/res/values/strings.xml b/app/src/main/res/values/strings.xml index 0fc0cb8c5a..996c81fb6b 100644 --- a/app/src/main/res/values/strings.xml +++ b/app/src/main/res/values/strings.xml @@ -2162,6 +2162,9 @@ Ad start marked. Press again at the end. Marking cancelled Cannot report an ad at this position + Rule saved and active + Nothing to save for this interval + Could not save the rule AI/custom %1$d · source config %2$d · built-in HLS %3$d Built-in HLS · v%1$d · %2$s Enabled diff --git a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index 7bee81b191..840e037f9e 100644 --- a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -65,8 +65,11 @@ import com.fongmi.android.tv.api.DanmakuApi; import com.fongmi.android.tv.api.SiteApi; import com.fongmi.android.tv.api.config.AdBlockStatsStore; import com.fongmi.android.tv.api.config.VodConfig; -import com.fongmi.android.tv.bean.AdDetectionRequest; -import com.fongmi.android.tv.bean.AdDetectionResult; +import com.fongmi.android.tv.ad.feedback.AdFeedbackController; +import com.fongmi.android.tv.ad.feedback.AdFeedbackHostAdapter; +import com.fongmi.android.tv.ad.feedback.AdFeedbackSession; +import com.fongmi.android.tv.ad.feedback.AdRulePlanApplier; +import com.fongmi.android.tv.ui.dialog.AdFeedbackDialog; import com.fongmi.android.tv.bean.AiConfig; import com.fongmi.android.tv.bean.CastVideo; import com.fongmi.android.tv.bean.Danmaku; @@ -103,7 +106,6 @@ import com.fongmi.android.tv.player.engine.PlaySpec; import com.fongmi.android.tv.player.lut.LutPreset; import com.fongmi.android.tv.player.lut.LutSetting; import com.fongmi.android.tv.player.lut.LutStore; -import com.fongmi.android.tv.service.AiAdDetectionService; import com.fongmi.android.tv.service.PlaybackService; import com.fongmi.android.tv.service.PersonalRecommendationService; import com.fongmi.android.tv.service.IntroSkipService; @@ -439,7 +441,11 @@ private int mAudioBackgroundRandomNonce; private TmdbItem mPendingTmdbSeasonChoice; private int mPersonalRecommendationGeneration; private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.recommendationExecutor()); - private int mAdFeedbackGeneration; + private AdFeedbackController mAdFeedback; + private AdFeedbackHostAdapter mAdFeedbackHost; + private AdFeedbackDialog mAdFeedbackDialog; + /** 标记模式下已打的起点，-1 表示未进入标记模式。 */ + private long mAdMarkStartMs = -1; private final List mTmdbMovedViews = new ArrayList<>(); private boolean pendingLutImport; private boolean skipPausePiP; @@ -1483,6 +1489,7 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco mBinding.control.action.danmaku.setOnClickListener(guarded(this::onDanmaku)); mBinding.control.action.adFeedback.setOnClickListener(view -> onAdFeedback()); + mBinding.control.action.adFeedback.setOnLongClickListener(view -> onAdFeedbackLongPress()); mBinding.control.action.episodes.setOnClickListener(view -> onEpisodes()); mBinding.control.action.text.setOnLongClickListener(view -> onTextLong()); mBinding.control.action.speed.setOnLongClickListener(view -> onSpeedLong()); @@ -3050,7 +3057,7 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco return; } if (!isAdFeedbackEnabled()) { - Notify.show(R.string.ad_feedback_ai_disabled); + Notify.show(R.string.ad_interval_invalid); return; } hideControl(); @@ -8659,15 +8666,10 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco } private boolean isAdFeedbackEnabled() { - // 功能开关 + 仅支持解析的格式(HLS/m3u8)才可反馈,因为去广分析依赖切片列表 - return Setting.isAiConfigReady() && Setting.isAdblock() && Setting.isAiAdDetection() && isAdFeedbackSupportedFormat(); - } - - private boolean isAdFeedbackSupportedFormat() { - if (player() == null) return false; - String url = player().getUrl(); - if (TextUtils.isEmpty(url)) return false; - return com.fongmi.android.tv.player.exo.MediaSourceFactory.isHlsUrl(url); + // AI 不再是硬门槛：本地归因通道（切片结构/域名/既有规则）不依赖 AI 配置。 + // 只要求可定位的时间轴 —— 直播和时长未知的内容无法框选区间。 + if (player() == null || TextUtils.isEmpty(player().getUrl())) return false; + return Setting.isAdblock() && !player().isLive() && player().getDuration() > 0; } private void setAdFeedbackVisible() { @@ -8675,78 +8677,139 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco applyActionButtonVisibility(); } + private AdFeedbackController adFeedback() { + if (mAdFeedback == null) { + mAdFeedbackHost = new AdFeedbackHostAdapter( + new AdFeedbackPlayback(), new AdFeedbackMetadata(), new AdFeedbackUi()); + mAdFeedback = new AdFeedbackController(mAdFeedbackHost); + } + return mAdFeedback; + } + + /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitAdFeedback() { - AdDetectionRequest request = buildAdDetectionRequest(); - if (request == null) { - Notify.show(R.string.ad_feedback_no_url); - return; + AdFeedbackController controller = adFeedback(); + AdBlockStatsStore.recordFeedback(getKey()); + AdFeedbackSession session; + if (mAdMarkStartMs >= 0) { + long start = mAdMarkStartMs; + mAdMarkStartMs = -1; + session = controller.onMarkedInterval(start, player().getPosition()); + } else { + session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } - // 记录 AI 反馈统计 - AdBlockStatsStore.recordFeedback(request.getSiteKey()); - - Notify.show(R.string.ad_feedback_analyzing); - int generation = ++mAdFeedbackGeneration; - AiConfig config = AiConfig.objectFrom(Setting.getAiConfig()); - Task.execute(() -> { - // Parse m3u8 evidence (blocking I/O) - enrichRequestWithM3u8Evidence(request); - AdDetectionResult result = new AiAdDetectionService(config).analyze(request); - runOnUiThread(() -> { - if (isFinishing() || isDestroyed() || generation != mAdFeedbackGeneration) return; - onAdDetectionResult(request, result); - }); - }); + if (session == null) Notify.show(R.string.ad_interval_invalid); } - private AdDetectionRequest buildAdDetectionRequest() { - if (player() == null || TextUtils.isEmpty(player().getUrl())) return null; - String url = player().getUrl(); - Uri uri = Uri.parse(url); - AdDetectionRequest request = new AdDetectionRequest(); + /** 长按：打起点进入标记模式，再次长按取消。 */ + private boolean onAdFeedbackLongPress() { + if (!isAdFeedbackEnabled()) return false; + if (mAdMarkStartMs >= 0) { + mAdMarkStartMs = -1; + Notify.show(R.string.ad_interval_mark_cancelled); + } else { + mAdMarkStartMs = player().getPosition(); + Notify.show(R.string.ad_interval_mark_start); + } + return true; + } + + private void showAdFeedbackSession(AdFeedbackSession session) { + if (isFinishing() || isDestroyed()) return; + if (mAdFeedbackDialog == null) mAdFeedbackDialog = new AdFeedbackDialog(this, this::applyAdRulePlan); + mAdFeedbackDialog.show(session); + } + + private void applyAdRulePlan(com.fongmi.android.tv.ad.feedback.AdAttribution plan) { Site site = getSite(); - request.setSiteKey(site == null ? getKey() : site.getKey()); - request.setSiteName(site == null ? "" : site.getName()); - request.setVodName(mHistory == null ? mBinding.name.getText().toString() : mHistory.getVodName()); - request.setFlagName(getFlag() == null ? "" : getFlag().getFlag()); - request.setEpisodeName(getEpisode() == null ? "" : getEpisode().getName()); - request.setUrlHost(uri.getHost()); - request.setUrlPath(uri.getPath()); - return request; - } - - private void enrichRequestWithM3u8Evidence(AdDetectionRequest request) { - if (player() == null || TextUtils.isEmpty(player().getUrl())) return; - String url = player().getUrl(); - if (!url.contains(".m3u8")) return; // Only parse m3u8 - try { - java.util.Map headers = player().getHeaders(); - com.fongmi.android.tv.bean.M3u8Evidence evidence = com.fongmi.android.tv.utils.M3u8Parser.parse(url, headers); - request.setEvidence(evidence); - } catch (Exception e) { - // Ignore parsing failures - } - } - - private void onAdDetectionResult(AdDetectionRequest request, AdDetectionResult result) { - // 记录 AI 分析结果统计 - AdBlockStatsStore.recordAiAnalysis(result != null && !result.isError()); - - if (result == null || result.isError()) { - Notify.show(result == null ? getString(R.string.ad_feedback_failed) : result.getErrorMessage()); - return; - } - if (result.isEmpty()) { - Notify.show(R.string.ad_feedback_no_ad); - return; - } - // Show preview dialog - com.fongmi.android.tv.ui.dialog.AdRulePreviewDialog.create(result).show(this, confirmedResult -> { - UserAdRule rule = UserAdRule.fromAiResult(confirmedResult, request.getSiteKey()); - com.fongmi.android.tv.api.config.UserAdRuleStore.add(rule); - Notify.show(R.string.ad_feedback_saved); + AdRulePlanApplier.Outcome outcome = AdRulePlanApplier.apply( + plan, site == null ? getKey() : site.getKey()); + Notify.show(switch (outcome) { + case APPLIED -> R.string.ad_interval_rule_saved; + case SKIPPED -> R.string.ad_interval_rule_skipped; + case FAILED -> R.string.ad_interval_rule_failed; }); } + private final class AdFeedbackPlayback implements AdFeedbackHostAdapter.Playback { + @Override + public long positionMs() { + return player() == null ? 0 : player().getPosition(); + } + + @Override + public long durationMs() { + return player() == null ? 0 : player().getDuration(); + } + + @Override + public String playUrl() { + return player() == null ? "" : player().getUrl(); + } + + @Override + public java.util.Map headers() { + return player() == null ? java.util.Map.of() : player().getHeaders(); + } + + @Override + public boolean hls() { + return player() != null + && com.fongmi.android.tv.player.exo.MediaSourceFactory.isHlsUrl(player().getUrl()); + } + + @Override + public boolean skipInterval(long startMs, long endMs, String feedbackId) { + return player() != null && player().skipUserAdInterval(startMs, endMs, feedbackId); + } + } + + private final class AdFeedbackMetadata implements AdFeedbackHostAdapter.Metadata { + @Override + public String siteKey() { + Site site = getSite(); + return site == null ? getKey() : site.getKey(); + } + + @Override + public String siteName() { + Site site = getSite(); + return site == null ? "" : site.getName(); + } + + @Override + public String vodName() { + return mHistory == null ? mBinding.name.getText().toString() : mHistory.getVodName(); + } + + @Override + public String flagName() { + return getFlag() == null ? "" : getFlag().getFlag(); + } + + @Override + public String episodeName() { + return getEpisode() == null ? "" : getEpisode().getName(); + } + } + + private final class AdFeedbackUi implements AdFeedbackHostAdapter.Ui { + @Override + public void runBackground(Runnable task) { + Task.execute(task); + } + + @Override + public void runOnUi(Runnable task) { + runOnUiThread(task); + } + + @Override + public void showSession(AdFeedbackSession session) { + showAdFeedbackSession(session); + } + } + private static class ShortDramaControlItem { private final View view; private final ViewGroup parent; diff --git a/app/src/testMobile/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivityLayoutTest.java b/app/src/testMobile/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivityLayoutTest.java index a6018e902d..464a4eb478 100644 --- a/app/src/testMobile/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivityLayoutTest.java +++ b/app/src/testMobile/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivityLayoutTest.java @@ -443,15 +443,26 @@ public class TmdbDetailActivityLayoutTest { assertTrue("wide and mobile inline controls must dispatch the ad feedback action", source.contains("binding.playerAdFeedback.setOnClickListener(guarded(this::onInlineAdFeedback));") && source.contains("detailActionView(R.id.adFeedback, View.class).setOnClickListener(guarded(this::onInlineAdFeedback));")); - assertTrue("detail inline playback must use the same HLS and AI availability gate as the native player", + assertTrue("detail inline ad feedback must only require a seekable VOD timeline, not AI config", source.contains("private boolean isInlineAdFeedbackEnabled()") - && source.contains("Setting.isAiConfigReady() && Setting.isAdblock() && Setting.isAiAdDetection()") - && source.contains("MediaSourceFactory.isHlsUrl(player().getUrl())")); - assertTrue("detail inline playback must submit AI analysis and save confirmed user rules", + && source.contains("Setting.isAdblock() && !player().isLive() && player().getDuration() > 0")); + assertFalse("AI must no longer gate the ad feedback button", + source.contains("Setting.isAiConfigReady() && Setting.isAdblock() && Setting.isAiAdDetection()")); + assertTrue("detail inline playback must route ad feedback through the shared controller", source.contains("private void submitInlineAdFeedback()") - && source.contains("new AiAdDetectionService(config).analyze(request)") - && source.contains("AdRulePreviewDialog.create(result).show(this, confirmedResult ->") - && source.contains("UserAdRuleStore.add(rule);")); + && source.contains("controller.onMarkedInterval(start, player().getPosition())") + && source.contains("controller.onQuickReport(mAdFeedbackHost.cachedEvidence())") + && source.contains("new AdFeedbackController(mAdFeedbackHost)")); + assertTrue("long press must arm the interval marking mode", + source.contains("private boolean onInlineAdFeedbackLongPress()") + && source.contains("binding.playerAdFeedback.setOnLongClickListener(view -> onInlineAdFeedbackLongPress());") + && source.contains("detailActionView(R.id.adFeedback, View.class).setOnLongClickListener(view -> onInlineAdFeedbackLongPress());")); + assertTrue("confirmed plans must be applied through the shared applier", + source.contains("AdRulePlanApplier.apply(") + && source.contains("case APPLIED -> R.string.ad_interval_rule_saved;")); + assertFalse("the duplicated per-activity AI request builder must be gone", + source.contains("private AdDetectionRequest buildInlineAdDetectionRequest()") + || source.contains("new AiAdDetectionService(config).analyze(request)")); } /** From 75859b5de42ae91b216a36167832bbbe10279417 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 03:15:52 +0800 Subject: [PATCH 14/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E7=94=9F?= =?UTF-8?q?=E6=88=90=E7=9A=84=20HLS=20=E8=A7=84=E5=88=99=E5=AE=9E=E9=99=85?= =?UTF-8?q?=E5=88=A0=E4=B8=8D=E6=8E=89=E5=B9=BF=E5=91=8A=E5=88=87=E7=89=87?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 端到端测试（GenRuleE2ETest）把分类器载荷真的组装成 HlsAdRule 交给 HlsManifestCleaner 跑 30 片 manifest，三个场景全部一片未删。单元测试 只验证了载荷字段与 compile() 通过，覆盖不到运行时评分。 根因是误解了 HlsManifestCleaner 的信号语义：它对每个切片独立累加信号 （HlsManifestCleaner.java:143-149），达到 minimumSignals 即删。因此只能 编码对块内每一片都成立、且有区分度的条件： - hostSuffixes 原本包含与 playlist 同域的域名，同域条件对每片正片都成立， 会误删正片。现只保留非本站域名。 - requireDiscontinuity 只对广告块首片成立，块内后续片拿不到该信号， 编码进规则会只删首片、留下其余广告。现不编码，断点只用于归因定位区间。 - 新增 pathPatterns：命中的广告目录段经 Pattern.quote 转正则，作为同域 场景下唯一的区分条件。两类区分条件都没有时返回空载荷让通道弃权。 - minimumSignals 从 Math.max(2, ...) 改为 Math.min(2, encoded)： 超过实际可编码信号数会让 compile() 抛 Invalid minimum signals， 规则永远写不进去且用户只看到「保存失败」。 - AdRulePlanApplier 不再二次抬高门限，同样原因。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback: 13 suites 117 tests 0 failures (GenRuleE2ETest 3 new) Task-Guard: AF5a-e2e-fix --- .../tv/ad/feedback/AdRulePlanApplier.java | 4 +- .../tv/ad/feedback/HlsSegmentClassifier.java | 67 ++++++++-- .../android/tv/ad/feedback/RulePayload.java | 12 +- .../tv/ad/feedback/GenRuleE2ETest.java | 120 ++++++++++++++++++ .../RulePayloadCompileContractTest.java | 2 +- .../tv/ad/feedback/RulePayloadTest.java | 8 +- 6 files changed, 198 insertions(+), 15 deletions(-) create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/GenRuleE2ETest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdRulePlanApplier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdRulePlanApplier.java index 94e6df8cf4..04a5eb1f7d 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdRulePlanApplier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdRulePlanApplier.java @@ -71,7 +71,9 @@ public final class AdRulePlanApplier { payload.hasDurationRange() ? payload.durationMin() : null, payload.hasDurationRange() ? payload.durationMax() : null, payload.requireDiscontinuity(), payload.requireCrossDomain(), - Math.max(2, payload.minimumSignals())); + // 门限由分类器按实际可编码信号数算好，不能在这里再抬高 —— + // 超过实际信号数会让 compile() 拒绝整条规则 + payload.minimumSignals()); return UserHlsRuleStore.add(rule) ? Outcome.APPLIED : Outcome.FAILED; } diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index 62c2eda652..07efbade05 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -5,6 +5,7 @@ import java.util.HashMap; import java.util.List; import java.util.Locale; import java.util.Map; +import java.util.regex.Pattern; /** * 基于 HLS 切片结构的归因通道。零新增运行时成本：全部输入来自 @@ -87,16 +88,28 @@ public final class HlsSegmentClassifier { /** * 生成可被 {@code HlsManifestCleaner.Rule} 执行的条件。 * - *

时长范围围绕区间内均值取 ±0.1s，仅在时长离群信号成立时给出 —— - * 否则会写出一条只靠固定时长删片的宽泛规则，这是内置规则策略明确禁止的。 - * minimumSignals 取可编码信号数且不低于 2。 + *

关键约束：{@code HlsManifestCleaner} 对每个切片独立累加信号， + * 达到 minimumSignals 即删除。因此只能编码**有区分度**的条件 —— + * 与 playlist 同域的 hostSuffixes 对每个切片都成立，requireDiscontinuity + * 对每个断点后的切片都成立，两者凑够 2 个信号会连正片一起删。 + * + *

有区分度的条件只有两类：非本站域名、广告路径特征。至少要有一类， + * 否则返回空载荷让通道弃权。 */ private static RulePayload payloadOf(AdIntervalEvidence evidence, Signals signals) { + String playlistHost = evidence.playlistHost(); + // 只保留非本站域名 —— 同域 host 不是区分条件 List hosts = evidence.inside().stream() .map(SegmentFact::host) .filter(host -> !host.isEmpty()) + .filter(host -> playlistHost.isEmpty() || !hostMatches(host, playlistHost)) .distinct() .toList(); + List pathPatterns = signals.pathHint() + ? pathPatternsOf(evidence.inside()) : List.of(); + // 没有任何区分条件时不得生成规则 + if (hosts.isEmpty() && pathPatterns.isEmpty()) return RulePayload.empty(); + double min = Double.NaN; double max = Double.NaN; if (signals.durationOutlier()) { @@ -104,12 +117,48 @@ public final class HlsSegmentClassifier { min = Math.max(0d, mean - 0.1d); max = mean + 0.1d; } - // 作用域必须收窄到当前站点的 playlist 域名，否则规则会污染其他站点 - List scope = evidence.playlistHost().isEmpty() - ? List.of() : List.of(evidence.playlistHost()); - return RulePayload.ofHlsRule(scope, hosts, min, max, - signals.discontinuity(), signals.crossDomain(), - Math.max(2, signals.encodableCount())); + // 作用域收窄到当前站点的 playlist 域名，否则规则会污染其他站点 + List scope = playlistHost.isEmpty() ? List.of() : List.of(playlistHost); + + // requireDiscontinuity 只对广告块首片成立，块内后续切片拿不到这个信号。 + // 把它计入门限会导致只删掉首片、留下其余广告，因此不编码进规则 —— + // 断点已经在归因阶段用于定位区间，规则层不再需要它。 + // requireCrossDomain 对块内每一片都成立，可以安全编码。 + boolean requireCrossDomain = signals.crossDomain(); + + // 逐项统计真正对块内每一片都成立的信号，minimumSignals 不能超过它， + // 否则 HlsAdRule.compile() 会拒绝整条规则。 + int encoded = 0; + if (!hosts.isEmpty()) encoded++; + if (!pathPatterns.isEmpty()) encoded++; + if (!Double.isNaN(min)) encoded++; + if (requireCrossDomain) encoded++; + + // 门限优先取 2 以避免宽泛删片，但不得超过实际可编码的信号数 + // （超过会让 compile() 拒绝整条规则）。也不取满全部信号 —— 要求 + // 三个条件同时成立时，广告块里稍有出入的一片就会漏删。 + int minimumSignals = Math.min(2, encoded); + + return RulePayload.ofHlsRule(scope, hosts, pathPatterns, min, max, + false, requireCrossDomain, minimumSignals); + } + + /** 命中的广告路径目录段转成正则，用 quote 避免元字符注入。 */ + private static List pathPatternsOf(List inside) { + List patterns = new ArrayList<>(); + for (String hint : PATH_HINTS) { + boolean allMatch = inside.stream() + .allMatch(fact -> fact.path().toLowerCase(Locale.US).contains(hint)); + if (allMatch) patterns.add(Pattern.quote(hint)); + } + return patterns; + } + + /** host 是否等于给定域名或为其子域。 */ + private static boolean hostMatches(String host, String domain) { + String lower = host.toLowerCase(Locale.US); + String target = domain.toLowerCase(Locale.US); + return lower.equals(target) || lower.endsWith("." + target); } /** 逐项检测五种信号。 */ diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RulePayload.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RulePayload.java index 12d8131fa6..b5db233817 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RulePayload.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RulePayload.java @@ -67,13 +67,23 @@ public record RulePayload(List playlistHostSuffixes, List hosts, /** HLS 结构化条件全量。 */ public static RulePayload ofHlsRule(List playlistHostSuffixes, List hosts, + List regex, double durationMin, double durationMax, boolean requireDiscontinuity, boolean requireCrossDomain, int minimumSignals) { - return new RulePayload(playlistHostSuffixes, hosts, List.of(), List.of(), "", + return new RulePayload(playlistHostSuffixes, hosts, regex, List.of(), "", durationMin, durationMax, requireDiscontinuity, requireCrossDomain, minimumSignals); } + /** HLS 结构化条件，无 URL 正则。 */ + public static RulePayload ofHlsRule(List playlistHostSuffixes, List hosts, + double durationMin, double durationMax, + boolean requireDiscontinuity, boolean requireCrossDomain, + int minimumSignals) { + return ofHlsRule(playlistHostSuffixes, hosts, List.of(), durationMin, durationMax, + requireDiscontinuity, requireCrossDomain, minimumSignals); + } + public boolean hasDurationRange() { return !Double.isNaN(durationMin) && !Double.isNaN(durationMax); } diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/GenRuleE2ETest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/GenRuleE2ETest.java new file mode 100644 index 0000000000..fe74c98b81 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/GenRuleE2ETest.java @@ -0,0 +1,120 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertNotNull; + +import com.fongmi.android.tv.bean.HlsAdRule; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +/** + * 端到端：分类器产出的载荷 → HlsAdRule → HlsManifestCleaner， + * 验证生成的规则在真实 manifest 上删对了切片、且不误删正片。 + */ +public class GenRuleE2ETest { + + private static final String BASE = "https://v.example.com/play/index.m3u8"; + private static final int TOTAL = 30; + private static final int AD_FROM = 10; + private static final int AD_COUNT = 3; + + private static List seg(int from, int n, String host, String path, + double dur, boolean discFirst) { + List l = new ArrayList<>(); + for (int i = 0; i < n; i++) { + l.add(new SegmentFact(from + i, host, path + (from + i) + ".ts", dur, discFirst && i == 0)); + } + return l; + } + + private static List outsideAround(String host, double dur) { + List l = new ArrayList<>(seg(0, AD_FROM, host, "/seg/", dur, false)); + l.addAll(seg(AD_FROM + AD_COUNT, TOTAL - AD_FROM - AD_COUNT, host, "/seg/", dur, true)); + return l; + } + + private static AdIntervalEvidence ev(List in, List out, + boolean bounded, boolean cross) { + return new AdIntervalEvidence(80_000, 100_000, StartOrigin.USER_MARKED, + "site", "站点", "剧", "线", "1", "v.example.com", "/play/index.m3u8", true, + in, out, bounded, cross, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + /** 按 AdRulePlanApplier 的方式组装规则。 */ + private static HlsManifestCleaner.Rule toRule(RulePayload p) { + HlsAdRule r = HlsAdRule.createUserRule("id", "n", + p.playlistHostSuffixes(), p.hosts(), p.regex(), + p.hasDurationRange() ? p.durationMin() : null, + p.hasDurationRange() ? p.durationMax() : null, + p.requireDiscontinuity(), p.requireCrossDomain(), + p.minimumSignals()); + return r.compile(); + } + + /** 30 片，AD_FROM 起 3 片是广告，广告块前后各有断点。 */ + private static String manifest(String adHost, String adPath, double adDur) { + StringBuilder b = new StringBuilder("#EXTM3U\n"); + for (int i = 0; i < TOTAL; i++) { + boolean ad = i >= AD_FROM && i < AD_FROM + AD_COUNT; + if (i == AD_FROM || i == AD_FROM + AD_COUNT) b.append("#EXT-X-DISCONTINUITY\n"); + b.append("#EXTINF:").append(ad ? adDur : 8.0).append(",\n"); + b.append(ad ? adHost + adPath + i + ".ts" : "https://v.example.com/seg/" + i + ".ts") + .append('\n'); + } + return b.append("#EXT-X-ENDLIST\n").toString(); + } + + @Test + public void crossDomainRuleRemovesExactlyTheAdBlock() { + AdAttribution plan = HlsSegmentClassifier.classify( + ev(seg(AD_FROM, AD_COUNT, "ad-cdn.other.com", "/seg/", 6.4, true), + outsideAround("v.example.com", 8.0), true, true)); + assertNotNull(plan); + + HlsManifestCleaner.Result res = HlsManifestCleaner.clean(BASE, + manifest("https://ad-cdn.other.com", "/seg/", 6.4), + List.of(toRule(plan.payload()))); + + assertEquals("只删广告块", AD_COUNT, res.removedSegments()); + } + + @Test + public void domainChannelRuleRemovesExactlyTheAdBlock() { + AdAttribution plan = DomainReputationClassifier.classify( + ev(seg(AD_FROM, AD_COUNT, "ad-cdn.other.com", "/seg/", 6.4, false), + outsideAround("v.example.com", 8.0), false, true), + new DomainReputationClassifier.Input( + List.of(), List.of("v.example.com"), List.of(), "")); + assertNotNull(plan); + + HlsManifestCleaner.Result res = HlsManifestCleaner.clean(BASE, + manifest("https://ad-cdn.other.com", "/seg/", 6.4), + List.of(toRule(plan.payload()))); + + assertEquals("只删广告块", AD_COUNT, res.removedSegments()); + } + + /** + * 同域广告块：广告与正片同域名、同时长，只有路径特征和断点可用。 + * 这是最危险的场景 —— 若规则把 playlist 域名当作切片域名条件， + * 会连带删掉每一个紧跟断点的正片切片。 + */ + @Test + public void sameDomainRuleMustNotDeleteLegitSegmentsAfterDiscontinuity() { + AdAttribution plan = HlsSegmentClassifier.classify( + ev(seg(AD_FROM, AD_COUNT, "v.example.com", "/ads/", 8.0, true), + outsideAround("v.example.com", 8.0), true, false)); + if (plan == null) return; // 通道弃权也是可接受的安全结果 + + HlsManifestCleaner.Result res = HlsManifestCleaner.clean(BASE, + manifest("https://v.example.com", "/ads/", 8.0), + List.of(toRule(plan.payload()))); + + assertEquals("只能删广告块，不得误删断点后的正片", AD_COUNT, res.removedSegments()); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadCompileContractTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadCompileContractTest.java index fd4093c95f..8b1c227798 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadCompileContractTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadCompileContractTest.java @@ -30,7 +30,7 @@ public class RulePayloadCompileContractTest { payload.hasDurationRange() ? payload.durationMin() : null, payload.hasDurationRange() ? payload.durationMax() : null, payload.requireDiscontinuity(), payload.requireCrossDomain(), - Math.max(2, payload.minimumSignals())); + payload.minimumSignals()); try { assertNotNull(label, rule.compile()); } catch (RuntimeException e) { diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java index adfe99a69b..741efc6a7c 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java @@ -89,10 +89,12 @@ public class RulePayloadTest { // 作用域收窄到本站，避免规则污染其他站点 assertEquals(List.of("v.example.com"), payload.playlistHostSuffixes()); assertEquals(List.of("ad-cdn.other.com"), payload.hosts()); - assertTrue(payload.requireDiscontinuity()); + // requireDiscontinuity 只对广告块首片成立，编码进规则会导致只删首片、 + // 留下其余广告，因此刻意不编码 —— 断点只在归因阶段用于定位区间 + assertFalse(payload.requireDiscontinuity()); assertTrue(payload.requireCrossDomain()); - // 永不生成 minimumSignals=1 的宽泛规则 - assertTrue(payload.minimumSignals() >= 2); + // 跨域 + 域名 + 时长离群三个信号，门限取 2 + assertEquals(2, payload.minimumSignals()); assertTrue(plan.actionable()); } From b8ca2b361586792e733ad3e062920ea3d38812d3 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 03:40:35 +0800 Subject: [PATCH 15/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E8=AF=84?= =?UTF-8?q?=E5=AE=A1=E5=8F=91=E7=8E=B0=E7=9A=84=E5=85=AD=E4=B8=AA=E9=97=AE?= =?UTF-8?q?=E9=A2=98=EF=BC=8C=E5=85=B6=E4=B8=AD=E4=B8=A4=E5=A4=84=E4=BD=BF?= =?UTF-8?q?=E5=8A=9F=E8=83=BD=E6=88=90=E4=B8=BA=E6=AD=BB=E4=BB=A3=E7=A0=81?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 1. 三端 submitAdFeedback 直接调 player().getPosition() 无 null 保护。 player() 在 service 断连时返回 null，PlayerManager.getPosition() 内部 也无守卫。改为经 AdFeedbackHostAdapter.safePositionMs() 取值，取不到 返回 -1；且先取位置再清 mAdMarkStartMs —— 原顺序会在异常时丢失标记。 2. AdFeedbackController.generation 非 volatile 但跨线程读写，加 volatile。 3. 换源换集不再作废在途归因。TmdbDetail 原来递增 mAdFeedbackGeneration 作废旧回调，接线时替换成 invalidateEvidence() 却丢了作废语义：后台 fetchEvidence() 返回后仍会用上一集的证据弹窗，用户按「保存规则」写入的 是上一集的域名与切片条件。新增 AdFeedbackController.invalidate()，三端 统一经 resetAdFeedback() 调用。mobile/leanback 原本连 invalidateEvidence() 都没接，换集后短按会拿上一集的切片结构推断起点。 4. recordPlaybackHost() 全仓无生产调用，导致 siteBaselineHosts() 恒为空、 DomainReputationClassifier 因「无基线不断言」而永久弃权 —— 整条域名通道 是死代码。三端在播放地址确定后接线。 5. AdFeedbackDialog 未在 onDestroy 关闭。usableActivity() 已挡住销毁后弹窗， 但 AlertDialog 强引用字段仍会经 dialog→Context 泄漏 Activity。三端 onDestroy 加 close()，与 AdSkipPromptPresenter 的既有做法一致。 6. AdSkipCoordinator.onUserInterval 的 seek 失败恢复复查无字段污染 （activeKey/activeCandidate/undoContext 在该方法内从未被写），保持现状。 TmdbDetailActivityLayoutTest 补四条断言锁定上述接线，防止回归。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback + ui.activity: 471 tests 0 failures; full suite 4013 tests with the same 5 pre-existing failures as baseline; compileMobileArm64_v8aDebug + compileLeanbackArm64_v8aDebug OK Task-Guard: AF5b-review-fixes --- .../android/tv/ui/activity/VideoActivity.java | 23 ++++++++++++- .../tv/ad/feedback/AdFeedbackController.java | 23 +++++++++++-- .../tv/ad/feedback/AdFeedbackHostAdapter.java | 32 ++++++++++++++++--- .../tv/ui/activity/TmdbDetailActivity.java | 27 +++++++++++++--- .../android/tv/ui/activity/VideoActivity.java | 23 ++++++++++++- .../TmdbDetailActivityLayoutTest.java | 17 +++++++++- 6 files changed, 130 insertions(+), 15 deletions(-) diff --git a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index 4fe1295f3a..b8d753bc35 100644 --- a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -2570,6 +2570,7 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; } private void getPlayer(Flag flag, Episode episode) { + resetAdFeedback(); mBinding.widget.title.setText(getPlaybackControlTitle(episode)); playerStartTime = System.currentTimeMillis(); beginPlayHealth(); @@ -4020,15 +4021,33 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; return mAdFeedback; } + /** 换源换集：作废在途归因、清标记模式与上一集的切片证据缓存。 */ + private void resetAdFeedback() { + mAdMarkStartMs = -1; + if (mAdFeedback != null) mAdFeedback.invalidate(); + if (mAdFeedbackHost != null) mAdFeedbackHost.invalidateEvidence(); + } + + /** 起播后记录本站域名，供域名信誉通道建立基线。 */ + private void recordAdFeedbackHost() { + if (mAdFeedbackHost != null) mAdFeedbackHost.recordPlaybackHost(); + } + /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitAdFeedback() { AdFeedbackController controller = adFeedback(); AdBlockStatsStore.recordFeedback(getKey()); AdFeedbackSession session; if (mAdMarkStartMs >= 0) { + // 先取位置再清标记：播放器已释放时保留标记，用户可重试而不必重新打点 + long end = mAdFeedbackHost.safePositionMs(); + if (end < 0) { + Notify.show(R.string.ad_interval_invalid); + return; + } long start = mAdMarkStartMs; mAdMarkStartMs = -1; - session = controller.onMarkedInterval(start, player().getPosition()); + session = controller.onMarkedInterval(start, end); } else { session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } @@ -4997,6 +5016,7 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; requestIntroSkipPlan(); if (!pendingResumeSeekApplied) applyAutoIntroSkip(); setAdFeedbackVisible(); // 播放地址确定后按格式刷新"有广告"按钮 + recordAdFeedbackHost(); // 播放地址已确定，记入本站域名基线 break; case Player.STATE_ENDED: checkEnded(true); @@ -7231,6 +7251,7 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; } @Override protected void onDestroy() { + if (mAdFeedbackDialog != null) mAdFeedbackDialog.close(); cancelAiSeasonAnalysis(false); mLyricsSearchSeq++; mLyricsRefreshSeq++; diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java index 98b160adcb..add0e942ed 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java @@ -74,8 +74,11 @@ public final class AdFeedbackController { private final Host host; private final AdFeedbackDiagnostics diagnostics; - /** 防止过期回调覆盖新一次反馈的结果。 */ - private int generation; + /** + * 防止过期回调覆盖新一次反馈的结果。写在主线程（{@link #submit}）、 + * 读在后台线程投递回主线程的 lambda 里，故必须 volatile。 + */ + private volatile int generation; public AdFeedbackController(Host host) { this(host, new AdFeedbackDiagnostics()); @@ -91,6 +94,15 @@ public final class AdFeedbackController { return diagnostics; } + /** + * 作废所有在途归因。换源、换集或播放器重建时必须调用 —— 否则后台 + * {@code fetchEvidence()} 返回后仍会用上一集的证据弹窗，用户按「保存规则」 + * 写入的是上一集的域名与切片条件。 + */ + public void invalidate() { + generation++; + } + /** * 短按：以当前位置为终点，回溯推断起点。 * @@ -104,7 +116,12 @@ public final class AdFeedbackController { return submit(start.startMs(), endMs, start.origin()); } - /** 长按标记模式：用户显式框选的区间。 */ + /** + * 长按标记模式：用户显式框选的区间。 + * + * @param startMs 标记模式下打的起点 + * @param endMs 终点，由调用方取当前播放位置；取不到时传负数即被拒绝 + */ public AdFeedbackSession onMarkedInterval(long startMs, long endMs) { return submit(startMs, endMs, StartOrigin.USER_MARKED); } diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java index 7fd1723534..abb4c509b9 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java @@ -108,16 +108,38 @@ public final class AdFeedbackHostAdapter implements AdFeedbackController.Host { cachedEvidence = null; } - /** 记录一次成功播放的域名，喂给站点基线。 */ + /** + * 当前播放位置，取不到时返回 -1 而非 0 —— 0 会被误当成合法的 + * 「片头位置」，而 -1 会让 {@code AdFeedbackController.submit} 直接拒绝。 + */ + public long safePositionMs() { + try { + long position = playback.positionMs(); + return position > 0 ? position : -1L; + } catch (RuntimeException e) { + return -1L; + } + } + + /** + * 记录一次成功播放的域名，喂给站点基线。 + * + *

必须在播放起播后调用：{@link DomainReputationClassifier} 无基线数据时 + * 一律弃权（避免首次播放即误判），不接线会让整条域名通道成为死代码。 + */ public void recordPlaybackHost() { - String host = AdEvidenceCollector.hostOf(playback.playUrl()); - if (host.isEmpty()) return; - deps.baseline().record(metadata.siteKey(), host); + try { + String host = AdEvidenceCollector.hostOf(playback.playUrl()); + if (host.isEmpty()) return; + deps.baseline().record(metadata.siteKey(), host); + } catch (RuntimeException ignored) { + // 基线只是归因的辅助输入，记录失败不应影响播放 + } } @Override public long positionMs() { - return playback.positionMs(); + return safePositionMs(); } @Override diff --git a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java index 65c1be4d59..62e8a08a21 100644 --- a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java +++ b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java @@ -6894,9 +6894,8 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. } private void stopInlinePlayerForReload() { - // 换源换集：作废标记模式并清掉上一集的切片证据缓存 - mAdMarkStartMs = -1; - if (mAdFeedbackHost != null) mAdFeedbackHost.invalidateEvidence(); + // 换源换集：作废在途归因，避免用上一集的证据弹窗并写错规则 + resetAdFeedback(); subtitlePlaybackSession.stop(this); inlineStartPosition = C.TIME_UNSET; inlineStartPositionApplied = false; @@ -8322,15 +8321,33 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. return mAdFeedback; } + /** 换源换集：作废在途归因、清标记模式与上一集的切片证据缓存。 */ + private void resetAdFeedback() { + mAdMarkStartMs = -1; + if (mAdFeedback != null) mAdFeedback.invalidate(); + if (mAdFeedbackHost != null) mAdFeedbackHost.invalidateEvidence(); + } + + /** 起播后记录本站域名，供域名信誉通道建立基线。 */ + private void recordAdFeedbackHost() { + if (mAdFeedbackHost != null) mAdFeedbackHost.recordPlaybackHost(); + } + /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitInlineAdFeedback() { AdFeedbackController controller = adFeedback(); AdBlockStatsStore.recordFeedback(getKeyText()); AdFeedbackSession session; if (mAdMarkStartMs >= 0) { + // 先取位置再清标记：播放器已释放时保留标记，用户可重试而不必重新打点 + long end = mAdFeedbackHost.safePositionMs(); + if (end < 0) { + Notify.show(R.string.ad_interval_invalid); + return; + } long start = mAdMarkStartMs; mAdMarkStartMs = -1; - session = controller.onMarkedInterval(start, player().getPosition()); + session = controller.onMarkedInterval(start, end); } else { session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } @@ -10270,6 +10287,7 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. player().reset(); applyInlineStartPosition(); updateInlineButtons(player().isPlaying()); + recordAdFeedbackHost(); // 播放地址已确定，记入本站域名基线 applyInlineShortDramaMode(); requestIntroSkipPlan(); applyAutoIntroSkip(); @@ -10430,6 +10448,7 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. @Override protected void onDestroy() { + if (mAdFeedbackDialog != null) mAdFeedbackDialog.close(); loadGeneration++; cancelAiSeasonAnalysis(false); detailTasks.close(); diff --git a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index 48f96ee2ce..1931b78dd9 100644 --- a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -2290,6 +2290,7 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco } private void getPlayer(Flag flag, Episode episode) { + resetAdFeedback(); mBinding.control.title.setText(getPlaybackControlTitle(episode)); playerStartTime = System.currentTimeMillis(); beginPlayHealth(); @@ -6721,6 +6722,7 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco requestIntroSkipPlan(); if (!pendingResumeSeekApplied) applyAutoIntroSkip(); setAdFeedbackVisible(); // 播放地址确定后按格式刷新"有广告"按钮 + recordAdFeedbackHost(); // 播放地址已确定，记入本站域名基线 break; case Player.STATE_ENDED: checkEnded(true); @@ -8656,6 +8658,7 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco } @Override protected void onDestroy() { + if (mAdFeedbackDialog != null) mAdFeedbackDialog.close(); cancelAiSeasonAnalysis(false); dismissKaraokeResultDialogForRecreation(); mLyricsSearchSeq++; @@ -8749,15 +8752,33 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco return mAdFeedback; } + /** 换源换集：作废在途归因、清标记模式与上一集的切片证据缓存。 */ + private void resetAdFeedback() { + mAdMarkStartMs = -1; + if (mAdFeedback != null) mAdFeedback.invalidate(); + if (mAdFeedbackHost != null) mAdFeedbackHost.invalidateEvidence(); + } + + /** 起播后记录本站域名，供域名信誉通道建立基线。 */ + private void recordAdFeedbackHost() { + if (mAdFeedbackHost != null) mAdFeedbackHost.recordPlaybackHost(); + } + /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitAdFeedback() { AdFeedbackController controller = adFeedback(); AdBlockStatsStore.recordFeedback(getKey()); AdFeedbackSession session; if (mAdMarkStartMs >= 0) { + // 先取位置再清标记：播放器已释放时保留标记，用户可重试而不必重新打点 + long end = mAdFeedbackHost.safePositionMs(); + if (end < 0) { + Notify.show(R.string.ad_interval_invalid); + return; + } long start = mAdMarkStartMs; mAdMarkStartMs = -1; - session = controller.onMarkedInterval(start, player().getPosition()); + session = controller.onMarkedInterval(start, end); } else { session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } diff --git a/app/src/testMobile/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivityLayoutTest.java b/app/src/testMobile/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivityLayoutTest.java index 464a4eb478..61ec91b5e2 100644 --- a/app/src/testMobile/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivityLayoutTest.java +++ b/app/src/testMobile/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivityLayoutTest.java @@ -450,9 +450,24 @@ public class TmdbDetailActivityLayoutTest { source.contains("Setting.isAiConfigReady() && Setting.isAdblock() && Setting.isAiAdDetection()")); assertTrue("detail inline playback must route ad feedback through the shared controller", source.contains("private void submitInlineAdFeedback()") - && source.contains("controller.onMarkedInterval(start, player().getPosition())") + && source.contains("controller.onMarkedInterval(start, end)") && source.contains("controller.onQuickReport(mAdFeedbackHost.cachedEvidence())") && source.contains("new AdFeedbackController(mAdFeedbackHost)")); + assertTrue("marked interval must read the position before clearing the mark so a released player keeps it", + source.contains("long end = mAdFeedbackHost.safePositionMs();") + && source.indexOf("long end = mAdFeedbackHost.safePositionMs();") + < source.indexOf("long start = mAdMarkStartMs;")); + assertTrue("source or episode switch must invalidate in-flight attribution", + source.contains("private void resetAdFeedback()") + && source.contains("mAdFeedback.invalidate();") + && source.contains("mAdFeedbackHost.invalidateEvidence();") + && source.contains("resetAdFeedback();")); + assertTrue("playback host must feed the site baseline or the domain channel stays dead", + source.contains("private void recordAdFeedbackHost()") + && source.contains("mAdFeedbackHost.recordPlaybackHost();") + && source.contains("recordAdFeedbackHost(); // 播放地址已确定，记入本站域名基线")); + assertTrue("the feedback dialog must be closed on destroy to avoid leaking the activity", + source.contains("if (mAdFeedbackDialog != null) mAdFeedbackDialog.close();")); assertTrue("long press must arm the interval marking mode", source.contains("private boolean onInlineAdFeedbackLongPress()") && source.contains("binding.playerAdFeedback.setOnLongClickListener(view -> onInlineAdFeedbackLongPress());") From e881bc9115354cddddfc9113f2fd9d322c37c469 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 04:15:28 +0800 Subject: [PATCH 16/28] =?UTF-8?q?=E6=B5=8B=E8=AF=95=EF=BC=9A=E9=94=81?= =?UTF-8?q?=E5=AE=9A=E7=89=87=E5=A4=B4=E9=92=B3=E4=BD=8D=E3=80=81=E4=BD=8D?= =?UTF-8?q?=E7=BD=AE=E4=B8=8D=E5=8F=AF=E7=94=A8=E4=B8=8E=E6=8D=A2=E9=9B=86?= =?UTF-8?q?=E4=BD=9C=E5=BA=9F=E4=B8=89=E4=B8=AA=E8=BE=B9=E7=95=8C?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 片头广告是最常见场景，位置远小于 90 秒回溯窗口，起点必须钳到 0； 之前只测过 200 秒位置的常规路径。 - safePositionMs() 返回 -1 时必须被拒，不能当成合法的片头位置 0。 - invalidate() 后过期归因回来不得再刷新界面 —— 这是上一个提交修的 换集串台问题，补测试防回归。 Verification: testMobileArm64_v8aDebugUnitTest AdFeedbackControllerTest: 16 tests 0 failures Task-Guard: AF5c-preroll-test --- .../ad/feedback/AdFeedbackControllerTest.java | 44 +++++++++++++++++++ 1 file changed, 44 insertions(+) diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java index 86785af554..34d56a1899 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java @@ -122,6 +122,50 @@ public class AdFeedbackControllerTest { .count(AdFeedbackDiagnostics.Code.START_INFERRED_FROM_FALLBACK)); } + @Test + public void prerollReportEarlyInPlaybackClampsStartToZero() { + // 片头广告是最常见场景：位置远小于回溯窗口，起点必须钳到 0 而非变负 + FakeHost host = new FakeHost(); + host.position = 12_000; + AdFeedbackController controller = new AdFeedbackController(host); + + AdFeedbackSession session = controller.onQuickReport(null); + + assertNotNull("片头位置也必须能反馈", session); + assertEquals(0L, session.startMs()); + assertEquals(12_000L, session.endMs()); + assertEquals(List.of("0-12000"), host.skipCalls); + } + + @Test + public void unavailablePositionIsRejectedNotTreatedAsZero() { + // 播放器已释放时 safePositionMs 返回 -1，不能被当成合法的片头位置 + FakeHost host = new FakeHost(); + host.position = -1; + AdFeedbackController controller = new AdFeedbackController(host); + + assertNull(controller.onQuickReport(null)); + assertTrue(host.skipCalls.isEmpty()); + assertTrue(host.shownSessions.isEmpty()); + } + + @Test + public void invalidateDropsInFlightAnalysisFromPreviousEpisode() { + // 换集后旧归因回来时不得覆盖 UI —— 否则用户会看到上一集的结论 + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + host.deferBackground = true; + AdFeedbackController controller = new AdFeedbackController(host); + + controller.onMarkedInterval(30_000, 60_000); + int shownBefore = host.shownSessions.size(); + controller.invalidate(); + host.deferBackground = false; + host.backgroundTasks.get(0).run(); + + assertEquals("过期归因不得再刷新界面", shownBefore, host.shownSessions.size()); + } + @Test public void rejectsInvalidIntervals() { FakeHost host = new FakeHost(); From 756954e2c5e26eaf9d337b3b137dc8de33c74f21 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 04:17:48 +0800 Subject: [PATCH 17/28] =?UTF-8?q?=E6=96=87=E6=A1=A3=EF=BC=9A=E5=8C=BA?= =?UTF-8?q?=E9=97=B4=E5=8F=8D=E9=A6=88=E4=B8=8E=E5=A4=9A=E9=80=9A=E9=81=93?= =?UTF-8?q?=E5=BD=92=E5=9B=A0=E8=AE=BE=E8=AE=A1?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 本次功能的设计依据，含背景、数据契约、各通道判据、仲裁规则、 分阶段计划与架构决策摘要。实现过程中修正的两处设计缺口也已回写： HlsAdRule 不支持位置信号、域名黑名单只在 WebView 层生效。 --- docs/ad-feedback-timerange-design.md | 527 +++++++++++++++++++++++++++ 1 file changed, 527 insertions(+) create mode 100644 docs/ad-feedback-timerange-design.md diff --git a/docs/ad-feedback-timerange-design.md b/docs/ad-feedback-timerange-design.md new file mode 100644 index 0000000000..fe9294f6bb --- /dev/null +++ b/docs/ad-feedback-timerange-design.md @@ -0,0 +1,527 @@ +# 「有广告」按钮升级：区间反馈与多通道综合归因设计 + +## 1. 文档状态 + +- 状态：提议（Proposed），待用户批准分阶段实施 +- 版本：v0.1 +- 日期：2026-08-29 +- 适用范围：播放器「有广告」按钮、广告归因分类器、候选规则落地、本次播放即时跳过 +- 上游文档：`智能去广-设计文档.md`（V2 总设计）、`docs/AI_AD_DETECTION_DESIGN.md`、`docs/audio-fingerprint-phase0-implementation.md`、`docs/interface-ad-rule-learning-design.md` +- 已确认决策：候选落地采用「本次立即跳过 + 同时生成待审候选」；通道分阶段接入；规则纯本地，不上传、不众包 + +## 2. 背景 + +### 2.1 现状 + +「有广告」按钮当前是**单通道**的：点击后只做一件事 —— 把站点元数据 + m3u8 切片摘要送给用户自配的 AI 端点，让大模型猜 `hosts` / `regex` / `exclude`，然后弹 `AdRulePreviewDialog` 让用户确认写入 `UserAdRule`。 + +三处入口（`app/src/leanback/.../VideoActivity.java:3982`、`app/src/mobile/.../VideoActivity.java:8661`、`app/src/main/.../TmdbDetailActivity.java:8286`）是三份几乎逐字重复的实现。 + +显示门槛是四个条件的与： +```java +Setting.isAiConfigReady() && Setting.isAdblock() + && Setting.isAiAdDetection() && isAdFeedbackSupportedFormat() +``` +其中 `isAiAdDetection()` 默认关闭（`Setting.java:868`），`isAdFeedbackSupportedFormat()` 要求 `MediaSourceFactory.isHlsUrl(url)`。**即按钮默认不可见，且强绑 AI。** + +### 2.2 三个具体缺陷 + +1. **反馈粒度只有「这一集有广告」**。用户明明知道广告在 02:15–02:45，这个信息被完整丢弃。分类器只能拿到整条 playlist 的统计特征，无法定位。 +2. **项目已有的其他检测能力完全没有参与**。仓库里已落地 HLS 结构化净化（`HlsManifestCleaner`，318 行）、音频频谱指纹（`ad/audio/` 41 个类，8000+ 行）、语音关键词（`SpeechAdSignalProvider`）、域名黑名单（`RuleConfig`）、接口规则学习（`InterfaceAdRuleAnalyzer`）—— 但「有广告」按钮一个都不用，只调 AI。 +3. **用户反馈完了什么都没得到**。当前流程最好的结果是「下次播放可能不再有广告」。本次播放的广告仍然要看完。 + +### 2.3 一个必须先修的既有问题 + +`HlsAdblockPipeline.apply()` 在两个调用点（`server/process/M3u8.java:91`、`androidx/media3/mpvplayer/MpvHlsProxy.java:799`）都传 `legacyFallback=true`。结构化规则未命中时会回落到预编译 AAR 里的 `HlsAdsParser.process()` —— 那是一套**没有规则、没有安全阈值、按众数块大小猜少数派**的启发式引擎。 + +由于内置的 3 条 HLS 规则全部 `enabledByDefault: false`（`assets/rules/hls_rules.json`），**当前线上实际生效的去广主力是这套启发式，而不是文档重点描述的结构化引擎。** 本设计的归因结论必须能区分「广告是被启发式删掉的」和「被结构化规则删掉的」，否则证据链是错的。 + +## 3. 目标与非目标 + +### 3.1 目标 + +1. 用户可提交广告的**开始/结束时间**，而不只是「有广告」这一个 bit。 +2. APP 用**已有的全部检测能力**对该区间做归因分类，输出「这段广告属于哪一类、最适合用哪种机制拦掉」。 +3. 分类结论按可靠性排序，选出**成本最低且证据最强**的那一种去广方式，而不是无脑走 AI。 +4. 用户提交后**当次播放立即跳过该区间**，不等分类完成。 +5. 分类产出的规则进入待审候选，用户确认后长期生效。 +6. 不引入服务端；不上传播放行为；AI 仍然可选且可完全关闭。 +7. 按钮在**关闭 AI 时也可用**（降级为纯本地归因）。 + +### 3.2 非目标 + +1. 不做视觉识别 / OCR / 帧分析。当前仓库无此能力，引入代价与收益不成比例。 +2. 不建服务端，不做跨用户众包聚合，不上报任何播放数据（承接现有「不承诺」清单）。 +3. 不为直播实现区间反馈（直播无稳定时间轴，`HlsManifestCleaner` 对 live 也只允许删连续头部分片）。 +4. 不承诺 IJK / MPV 的音频指纹通道（现有限制照旧，仅 Exo 支持 PCM 捕获）。 +5. 不自动启用高风险规则（宽泛固定时长、全文正则）。 +6. 不改动 `HlsAdsParser`（在预编译 AAR 内，属于 upstream 依赖，改它要走 `E*` 任务流程）。 + +## 4. 用户交互设计 + +### 4.1 区间采集：三种输入路径 + +单个按钮承担不了区间输入。改为**长按/短按分流 + 一个轻量标记条**。 + +| 操作 | 行为 | 适用 | +|---|---|---| +| 短按「有广告」 | 以当前位置为**结束时间**，向前回溯猜测开始时间（见 4.2），直接进入确认 | 广告刚放完，用户反应过来了。最常见 | +| 长按「有广告」 | 进入**标记模式**：第一次确认打起点，第二次确认打终点 | 广告正在放，用户想精确框选 | +| 标记模式中再次长按 | 取消标记 | 误触 | + +TV（leanback）端焦点资源紧张，长按走 `OnLongClickListener`；移动端同时在标记模式下于进度条上叠加区间高亮，允许拖动微调两个端点。 + +### 4.2 短按的起点推断 + +短按只有一个时间点。起点按优先级推断，第一个成功的即采用： + +1. **最近一次 `DISCONTINUITY` 边界**。从 `M3u8Evidence.discontinuities` 找当前位置之前最近的断点，落在其上。这是最可靠的信号 —— 硬插广告几乎总带断点。 +2. **最近一次跨域切换点**。`M3u8Evidence.domainSwitches` 中当前位置前最近的 `true`。 +3. **最近一次音频/语音候选的 `startTimeMs`**。若 `AdSkipCoordinator` 在 `SUPPRESSED` 或 `UNDO_WINDOW` 状态持有过候选，直接复用。 +4. **固定回溯窗口**。默认 90 秒，钳制到 `[0, 当前位置)`。 + +推断出的起点必须在确认对话框里**明示并可调**，不能静默使用。文案示例：「广告区间 02:15 – 02:45（起点由切片断点推断）」。 + +### 4.3 确认对话框 + +复用并扩展 `AdRulePreviewDialog`（leanback / mobile 各一份）。分两个阶段展示： + +``` +第一屏（提交即出，< 100ms） + 广告区间 02:15 – 02:45 (30.0s) [调整] [确认] + ☑ 本次播放立即跳过这段 + +第二屏（归因完成后原地替换内容，通常 1–3s，含 AI 时可能 10s+） + 归因结果：切片跨域 + 固定时长（置信度 0.82） + 建议方式：HLS 结构化规则 [风险：中] + hostSuffixes: ad-cdn.example.com + durationRange: 6.40 – 6.47 + requireDiscontinuity: true + minimumSignals: 2 + 证据：区间内 5 个切片全部来自 ad-cdn.example.com， + 主 playlist 域名为 v.example.com；区间前后各有一个 + #EXT-X-DISCONTINUITY；切片时长标准差 0.02s + [保存为规则] [仅本次跳过] [丢弃] +``` + +「本次播放立即跳过」默认勾选，勾选状态即视为对**本片段本次执行**的授权，与「保存为规则」的长期授权分离。这是决策 A 的落点。 + +### 4.4 按钮可见性放宽 + +```java +// 新条件：去广总开关 + 可定位的时间轴，AI 不再是硬门槛 +private boolean isAdFeedbackEnabled() { + return Setting.isAdblock() + && !isLive() + && getDuration() > 0 + && isSeekable(); +} +``` + +AI 从「按钮显示前置条件」降级为「归因通道之一」。关闭 AI 时按钮仍可用，只是少一个通道。HLS 格式也不再是硬门槛 —— 非 HLS 时 HLS 通道自然弃权，音频/语音通道仍可工作。 + +## 5. 总体架构 + +```mermaid +flowchart TD + U["用户提交 [startMs, endMs]"] --> R["AdFeedbackRequest 构造"] + R --> IMM["立即执行：本次跳过 (AdSkipCoordinator)"] + R --> COL["证据采集 AdEvidenceCollector"] + COL --> E["AdIntervalEvidence 快照"] + E --> C1["HlsSegmentClassifier"] + E --> C2["DomainReputationClassifier"] + E --> C3["ExistingRuleClassifier"] + E --> C4["AudioFingerprintClassifier"] + E --> C5["SpeechKeywordClassifier"] + E --> C6["AiClassifier (可选)"] + C1 --> A["AdAttributionArbiter 仲裁"] + C2 --> A + C3 --> A + C4 --> A + C5 --> A + C6 --> A + A --> P["AdRemediationPlan 最优方案"] + P --> UI["确认对话框第二屏"] + UI -->|用户确认| S["写入 UserAdRule / HlsAdRule / AudioFingerprintRule"] + S --> INV["RuleConfig.invalidate() + HlsRuleConfig.invalidate()"] +``` + +### 5.1 模块职责 + +| 模块 | 职责 | 不负责 | +|---|---|---| +| `AdFeedbackController` | 三端共用的编排器：接收区间、触发立即跳过、调度分类、驱动对话框 | 具体分类逻辑 | +| `AdEvidenceCollector` | 采集区间证据快照，一次性、幂等、可缓存 | 判断是不是广告 | +| `AdIntervalClassifier`（接口） | 单通道分类，输出 0..n 个 `AdAttribution` 或弃权 | 跨通道比较 | +| `AdAttributionArbiter` | 合并、去重、置信度融合、成本排序，选出最优方案 | 写规则 | +| `AdRemediationPlanner` | 把仲裁结论翻译成具体规则对象（`HlsAdRule` / `UserAdRule` / `AudioFingerprintRule`） | 落盘 | +| `AdFeedbackDiagnostics` | 固定枚举计数，不记 URL / 规则正文 | 上报 | + +### 5.2 三端去重 + +现有三份重复实现是这次改造的最大阻力。新逻辑**全部放进 `app/src/main/` 的 `AdFeedbackController`**，三端 Activity 只保留： + +```java +// 每端约 15 行 +mAdFeedback = new AdFeedbackController(new AdFeedbackController.Host() { + public long positionMs() { return mPlayer.getPosition(); } + public long durationMs() { return mPlayer.getDuration(); } + public String playUrl() { return mPlayer.getUrl(); } + public Map headers() { return mPlayer.getHeaders(); } + public AdSkipCoordinator skipCoordinator() { return mPlayer.adSkipCoordinator(); } + public void showDialog(AdFeedbackSession s) { AdRulePreviewDialog.create(s).show(this); } + public String vodName() { /* 三端唯一真正不同的地方 */ } +}); +``` + +现有 `testMobile` 下的 `PlayerPlaybackRegressionSourceTest` / `TmdbDetailActivityLayoutTest` 用**源码文本断言**锁定了三端的方法名，改动时必须同步更新这两个测试的期望字符串，否则构建失败。 + +## 6. 数据契约 + +### 6.1 区间证据 + +```java +public record AdIntervalEvidence( + long startMs, long endMs, StartOrigin startOrigin, + // 播放上下文（沿用 AdDetectionRequest 的去敏原则） + String siteKey, String siteName, String vodName, + String flagName, String episodeName, + String playlistHost, String urlPath, boolean hls, + // HLS 切片证据（区间内 + 区间外对照） + List inside, List outside, + boolean boundedByDiscontinuity, boolean crossDomain, + // 已有机制的自述 + List alreadyRemovedByStructuredRuleIds, + boolean legacyHeuristicActive, + List matchedExistingHosts, + // 音频/语音 + AudioIntervalFact audio, SpeechIntervalFact speech) {} + +public record SegmentFact(int index, String host, String path, + double durationSec, boolean discontinuityBefore) {} +``` + +`StartOrigin` 枚举：`USER_MARKED` / `DISCONTINUITY` / `CROSS_DOMAIN` / `AUDIO_CANDIDATE` / `FALLBACK_WINDOW`。仲裁时 `USER_MARKED` 与 `DISCONTINUITY` 的证据权重高于 `FALLBACK_WINDOW`。 + +`alreadyRemovedByStructuredRuleIds` 与 `legacyHeuristicActive` 直接回答 2.3 的问题：如果用户报告的区间**在净化后的 manifest 里已经不存在**，说明广告来自别处（例如 WebView 层没拦住的贴片、或播放器外层），归因结论应当是「HLS 通道弃权」而不是「再加一条 HLS 规则」。 + +### 6.2 归因与方案 + +```java +public record AdAttribution( + String channelId, // "hls" / "domain" / "existing-rule" / "audio" / "speech" / "ai" + AdCategory category, + float confidence, // 0..1，通道内自评 + RiskLevel risk, // LOW / MEDIUM / HIGH，沿用 V2 总设计的三级 + List evidence, // 人类可读，进 UI + RemediationKind remediation) {} + +public enum AdCategory { + THIRD_PARTY_CDN_SEGMENT, // 切片来自非主流/非当前适配域名 + FIXED_DURATION_BLOCK, // 固定时长硬插块 + DISCONTINUITY_BLOCK, // 断点包裹的独立块 + KNOWN_AD_HOST, // 命中已知广告域名 + AUDIO_FINGERPRINT, // 音频频谱指纹匹配 + SPEECH_KEYWORD, // 语音关键词 + IN_STREAM_BURNED_IN, // 压制进正片，无结构特征 + ALREADY_HANDLED, // 已被现有规则处理，用户看到的是别的东西 + UNKNOWN +} + +public enum RemediationKind { + HLS_STRUCTURED_RULE, // → HlsAdRule + HOST_BLACKLIST, // → UserAdRule.hosts + URL_REGEX_RULE, // → UserAdRule.regex + AUDIO_FINGERPRINT_RULE,// → AudioFingerprintRule + SPEECH_KEYWORD, // → SpeechAdSetting keywords + SESSION_SKIP_ONLY, // 只能本次跳过，无法泛化 + NONE +} +``` + +`AdRemediationPlan` = 首选 `AdAttribution` + 生成好的规则对象 + 备选方案列表。UI 展示首选，「更多方式」可展开备选。 + +## 7. 各通道分类逻辑 + +### 7.1 HlsSegmentClassifier（Phase 1，零新增成本） + +输入已有：`M3u8Parser.parse()` 已经产出 `segments` / `discontinuities` / `durations` / `domainSwitches`。缺的只是**把播放时间映射到切片下标** —— 用 `durations` 前缀和即可，不需要新的网络请求。 + +``` +cum[0]=0; cum[i+1]=cum[i]+durations[i] +区间内切片 = { i : cum[i+1] > startSec && cum[i] < endSec } +``` + +判定信号（沿用 `HlsManifestCleaner.Rule` 已支持的六项，保证生成的规则一定能被现有引擎执行）： + +| 信号 | 判据 | 权重 | +|---|---|---| +| 跨域 | 区间内切片 host ≠ playlist host，且区间外切片 host = playlist host | 0.35 | +| 断点包裹 | 区间起止各有 `#EXT-X-DISCONTINUITY` | 0.25 | +| 时长离群 | 区间内切片时长标准差 < 0.05s 且与区间外众数时长差 > 0.5s | 0.20 | +| 路径特征 | 区间内切片路径命中 `/ads?/`、`/preroll/`、`/creative/` 等保守模式 | 0.15 | +| 位置 | 区间落在 playlist 头部 15% 以内 | 0.05 | + +置信度 = 命中权重之和。`minimumSignals` 按命中信号数取 `max(2, 命中数)`，**永不生成 `minimumSignals: 1` 的规则** —— 内置规则策略明确禁止只凭固定时长删片。 + +生成的 `HlsAdRule` 必须带 `playlistHostSuffixes`（限定到当前站点的 playlist 域名），把规则作用域收窄，避免污染其他站点。 + +**弃权条件**：非 HLS；区间内切片数为 0；`alreadyRemovedByStructuredRuleIds` 非空且覆盖了整个区间；置信度 < 0.30。 + +### 7.2 DomainReputationClassifier（Phase 1，零新增成本） + +「非当前适配的主流域名」这条需求落在这里。三层比对： + +1. **命中现有黑名单**：区间内切片 host 命中 `RuleConfig.get().getAds()`（VOD `ads` + Live `ads` + 用户规则 hosts）→ `KNOWN_AD_HOST`，置信度 0.95，但同时标记 `ALREADY_HANDLED`：既然已在黑名单里，用户还是看到了广告，说明拦截路径没覆盖到播放器直连的切片请求（黑名单目前主要在 `CustomWebView.shouldInterceptRequest` 生效，不拦播放器的切片）。这个结论本身就有价值 —— 建议改用 `HLS_STRUCTURED_RULE`。 +2. **与站点适配域名对比**：从当前站点最近若干次成功播放的 playlist host 归纳出「本站正常域名集合」（新增一个小的本地 LRU，见 8.3）。区间内 host 不在其中 → `THIRD_PARTY_CDN_SEGMENT`。 +3. **命中接口学习候选**：比对 `ImportedAdRuleCandidateStore` 里状态为待审的候选。若命中，说明接口维护者也认为这是广告域名 → 置信度 +0.15，并在证据里注明来源接口名。这条把 `InterfaceAdRuleAnalyzer` 的既有产出接进了反馈闭环。 + +生成 `UserAdRule.hosts`。风险等级 LOW（域名黑名单是最安全的机制）。 + +### 7.3 ExistingRuleClassifier（Phase 1，零新增成本） + +回答「为什么现有机制没拦住」。检查四件事并写进证据，本身不产出规则（`RemediationKind.NONE`），但会**否决**其他通道的错误结论： + +1. 内置 3 条 HLS 规则是否因 `enabledByDefault: false` 而未启用，且其 `hostSuffixes` 恰好匹配本区间 → 建议「启用规则 X」而不是新建规则。这是成本最低的修复。 +2. `legacyHeuristicActive` 为真且区间内切片是被启发式删掉后**又出现**的 → 启发式判错，需要 exclude 保护。 +3. 是否有 `UserAdRule.exclude`（正片保护正则）误保护了广告切片。 +4. 区间是否已被某条规则处理过（查 `AdBlockStats.ruleCounts` 的最近命中）。 + +### 7.4 AudioFingerprintClassifier（Phase 2） + +Phase 1 只**读**已有匹配结果（若音频指纹已开启且区间内有 `MatchEvent`，直接采信，置信度 0.90，`RemediationKind.AUDIO_FINGERPRINT_RULE` 但规则已存在故降为证据）。 + +Phase 2 才做**新指纹采集**，这是真正的能力增量：用户框选的区间就是天然的指纹样本源。 + +``` +1. AdSkipCoordinator 在 UNDO_WINDOW / 提交瞬间，向 PlaybackMediaSignalHub + 申请 ConsumerKind.AD_AUDIO 的 CaptureLease（若尚未持有） +2. 若区间已过去 → 不重放，走「下次遇到再采集」：把区间描述存为 + PendingFingerprintCapture，下次同一 siteKey + 同一 flagName 播放 + 到相近位置时自动采集 +3. 若区间尚未播完（长按标记模式下常见）→ 直接从 Hub 的 PcmFrame 流采集 +4. SpectralFingerprint.extractVariants(samples, 16000, 1, ...) 产出 4 个相位变体 +5. 组装 AudioFingerprintRule{id, durationMs, anchorOffsetMs, anchorDurationMs, + fingerprint, variants}，经 AudioFingerprintRuleCodec 严格校验后 + 写入 AdAudioRuleStore（files/ad-audio-rules.json，2MiB 上限） +``` + +约束（全部继承自现有实现，不放松）： +- 仅 Exo；仅点播；仅时长已知且可 seek +- 采集不改变现有「必须用户确认才 seek」的策略 +- 锚点长度取区间前 3 秒（`anchorDurationMs` 上限），不存整段广告音频 +- **不存原始 PCM**，只存 32-bit hash 序列。这一点对隐私和体积都关键 +- 新指纹默认 `PROMPT` 模式，不自动跳 + +跨规则碰撞检查：新指纹与已有规则做汉明距离比对，过近则拒绝写入并提示「已有相同规则」，避免规则库膨胀。 + +### 7.5 SpeechKeywordClassifier（Phase 2） + +若语音通道已开启且区间内有识别文本，提取候选关键词。**但不自动加入关键词表** —— 关键词是全局生效的，一个错误关键词会在所有片子上乱跳。只做建议，且： +- 只建议在多个不同片源的反馈中重复出现的词（需要一个本地计数，阈值 3 次） +- 建议词展示时不显示完整识别文本（现有实现刻意不打印命中关键词，保持一致） + +### 7.6 AiClassifier（Phase 3，可选） + +复用 `AiAdDetectionService`，但输入升级： + +``` +现在只送：站点/剧名/线路/集名/域名/路径 + 全 playlist 切片摘要 +升级后送：上述 + 用户框选区间 + 区间内外切片对照 + 其他通道的初步结论 +``` + +关键变化是 **AI 从「唯一裁判」变成「兜底解释者」**。它只在以下情况被调用： +- 其他所有通道置信度均 < 0.50，或 +- 用户在第二屏主动点「让 AI 再看看」 + +理由：本地通道有确定性证据时，大模型的输出只会引入噪声和延迟；而本地通道全部弃权时（`IN_STREAM_BURNED_IN` 场景），AI 至少能给出人类可读的解释。 + +`AiConfig` 未配置时静默弃权，不报错。AI 输出的规则**永远进待审**，绝不自动启用（决策 D 不变）。 + +## 8. 仲裁与置信度 + +### 8.1 成本-可靠性排序 + +仲裁不是简单取最高置信度，而是在**同等置信度下优先选成本更低、风险更小的机制**： + +| 机制 | 运行时成本 | 风险 | 泛化能力 | 排序权重 | +|---|---|---|---|---| +| 启用已有内置规则 | 零 | 低 | 高 | 1（最优） | +| 域名黑名单 | 极低 | 低 | 中 | 2 | +| HLS 结构化规则 | 低（P95 < 50ms） | 中 | 高 | 3 | +| URL 正则规则 | 低 | 中 | 中 | 4 | +| 音频指纹 | 中（PCM 管线 + FFT） | 中 | 仅同一广告素材 | 5 | +| 语音关键词 | 高（Sherpa-ONNX 常驻） | 高（全局误伤） | 低 | 6 | +| 本次跳过 | 零 | 低 | 无 | 7（兜底） | + +最终得分 = `confidence × 0.7 + (1 - 归一化排序权重) × 0.3`。`ALREADY_HANDLED` 类别强制降权到最后。 + +### 8.2 冲突处理 + +- 多通道指向**同一** `AdCategory`：置信度取 `1 - Π(1 - cᵢ)`（概率或），证据合并 +- 多通道指向**不同** category：全部保留，按 8.1 排序，UI 展示首选 + 「更多方式」 +- 任一通道给出 `ALREADY_HANDLED`：整体降级为「诊断结论」，首选动作变成「本次跳过 + 上报诊断到本地日志」，不新增规则。避免规则库被无效规则污染 + +### 8.3 站点域名基线（新增最小存储） + +`SitePlaylistHostBaseline`，SharedPreferences，每站点保留最近 8 个成功播放的 playlist host + 切片 host，LRU 淘汰。仅用于 7.2 的第 2 层比对。 + +不存 URL 全文、不存 query、不存 token，只存 host 字符串。单站点上限 8，全局上限 200 站点，超限按最久未用淘汰。 + +## 9. 立即跳过的执行 + +这是「本次立即跳过」决策的落点。复用 `AdSkipCoordinator` 而不是新写 seek 逻辑，因为它已经处理了所有难的部分：session/generation 校验、时钟新鲜度、直播拒绝、不可 seek 拒绝、5 秒撤销窗口。 + +新增一条入口： + +```java +// AdSkipCoordinator 新增 +public synchronized boolean onUserInterval(long startMs, long endMs, String feedbackId); +``` + +与现有 `onCandidate` / `onAutoCandidate` 的区别： +- 用户已经显式授权，跳过 `PROMPT_PENDING` 直接进 `SEEKING` +- 但**仍然执行全部安全校验**：session/generation 匹配、`endMs < durationMs`、可 seek、非直播。任一失败则 `diagnostics.record(SEEK_REJECTED)` 并降级为「已记录，未跳过」 +- 保留 5 秒 `UNDO_WINDOW`。用户框错了要能立刻回去 +- 若 `endMs` 已过（短按场景，广告已放完）→ 不 seek，只记录，第一屏文案改为「已记录该区间」 + +**同一区间的重复提交**要幂等：以 `(sessionId, startMs/1000, endMs/1000)` 为键去重，避免用户连按导致多条候选。 + +## 10. 安全与隐私 + +继承现有边界，并针对本次新增的数据做收紧： + +1. **区间时间戳不出设备**。除了用户自配的 AI 端点（Phase 3，且用户显式开启），无任何网络出口。 +2. **AI 输入去敏**：只送 host + 去参数 path + 时长数组 + 断点下标。不送完整 URL、query、token、Cookie、Authorization。现有 `AdDetectionRequest` 已遵守此约定，扩展字段沿用。 +3. **`AiDebugLog` 的落盘风险需要处理**。现有 `AiAdDetectionService.java:44,50,58` 会把完整 prompt 与响应写本地日志。区间反馈会让 prompt 包含更多播放上下文。建议：debug 日志只在 debug 构建启用，或对 host 做哈希化。这是一个既有问题，本设计不扩大它。 +4. **音频指纹不存 PCM**，只存 hash 序列（7.4）。 +5. **正则安全**：任何生成的正则在写入前必须过 `InterfaceAdRuleAnalyzer` 已有的危险模式检查（嵌套量词、反向引用），并预编译验证。复用现有实现，不新写。 +6. **规则作用域强制收窄**：反馈生成的 `HlsAdRule` 必须带 `playlistHostSuffixes`，`UserAdRule` 必须带 `siteKey`。禁止生成全局生效的规则。 + +## 11. 可观测性 + +沿用 `AdAudioDiagnostics` 的「固定枚举计数，不记敏感内容」模式，新增 `AdFeedbackDiagnostics.Code`： + +``` +INTERVAL_SUBMITTED, START_INFERRED_FROM_DISCONTINUITY, +START_INFERRED_FROM_FALLBACK, IMMEDIATE_SKIP_APPLIED, +IMMEDIATE_SKIP_REJECTED, EVIDENCE_COLLECT_FAILED, +CHANNEL_ABSTAINED, ARBITER_NO_PLAN, PLAN_ACCEPTED, +PLAN_DISCARDED, DUPLICATE_SUBMISSION, ALREADY_HANDLED_DETECTED +``` + +`AdBlockStats` 扩展三个计数（沿用现有 SharedPreferences 存储，不建表）： +- `intervalFeedbackCount` — 区间反馈次数 +- `intervalSkipApplied` — 成功立即跳过次数 +- `planAcceptedByKind` — 按 `RemediationKind` 统计被采纳的方案分布 + +最后一项是**衡量本设计是否成功的核心指标**：如果一年后 `SESSION_SKIP_ONLY` 占比仍然最高，说明归因基本没用，应当回退简化。 + +## 12. 测试策略 + +### 12.1 单元测试（Phase 1 必须） + +- `AdEvidenceCollectorTest`：时间→切片下标映射，含边界（区间恰好落在切片边界、区间跨越 `EXT-X-DISCONTINUITY`、`durations` 缺失） +- `HlsSegmentClassifierTest`：五种信号的正样本各一、反样本各一；「区间内外同域名」必须弃权而非误判 +- `DomainReputationClassifierTest`：三层比对；`ALREADY_HANDLED` 的识别 +- `ExistingRuleClassifierTest`：内置规则未启用时建议启用而不是新建 +- `AdAttributionArbiterTest`：同 category 概率或、跨 category 排序、`ALREADY_HANDLED` 强制降权 +- `AdSkipCoordinatorTest` 扩展：`onUserInterval` 的 session/generation 拒绝路径、`endMs` 已过的降级、重复提交幂等 + +### 12.2 样本回归（Phase 1 必须） + +复用现有 `app/src/test/resources` 的 manifest 样本模式，为三条内置实验规则（暴风 / 量子 / 非凡）各准备一组：真实 manifest + 用户区间 + 期望归因结论。这些样本已有正/反样本，成本低。 + +### 12.3 Phase 2 追加 + +- `SpectralFingerprint` 黄金 PCM 向量端到端比对（`docs/ad-audio-fingerprint-sdk-evaluation.md` 4.1 节的既有待办，Phase 2 必须先关掉） +- 指纹碰撞检查：新采集指纹与已有规则过近时拒绝写入 +- `PendingFingerprintCapture` 的跨会话恢复 + +### 12.4 三端一致性 + +现有 `PlayerPlaybackRegressionSourceTest` / `TmdbDetailActivityLayoutTest` 靠源码文本断言锁定三端方法名。改造为 `AdFeedbackController` 后，这两个测试的期望字符串必须同步更新，且新增一条断言：三端**不得**各自持有 `submitAdFeedback` / `buildAdDetectionRequest` 的实现（防止重复实现回归）。 + +## 13. 分阶段实施 + +按「先零成本，再重成本」决策拆分。每个 Phase 独立可用、独立可回滚。 + +### Phase 1：区间输入 + 纯本地归因（无新增运行时成本） + +范围：`AdFeedbackController`、`AdEvidenceCollector`、`HlsSegmentClassifier`、`DomainReputationClassifier`、`ExistingRuleClassifier`、`AdAttributionArbiter`、`AdSkipCoordinator.onUserInterval`、对话框两屏、三端去重、`SitePlaylistHostBaseline`。 + +不含任何音频采集、不含 AI 改动（AI 保持现状可用）。 + +验收： +1. 长按可框选区间，短按可回溯推断起点且起点来源可见可调 +2. 提交后当次播放立即跳过（`endMs` 未过时），5 秒内可撤销 +3. 三条内置实验规则的样本 manifest 上，归因结论与人工标注一致 +4. 区间已被现有规则处理时，输出 `ALREADY_HANDLED` 而非新增规则 +5. 关闭 AI 时按钮可用且归因正常 +6. 三端不再存在重复的反馈实现 + +回滚：`AdFeedbackController` 的入口加一个 Setting 开关，关闭后按钮回到现有单通道 AI 行为。 + +### Phase 2：音频指纹采集闭环 + +范围：`AudioFingerprintClassifier` 的采集路径、`PendingFingerprintCapture`、指纹碰撞检查、`SpeechKeywordClassifier` 的建议（不自动启用）。 + +**前置条件**：黄金 PCM 向量比对完成（12.3）。未完成不得进入 Phase 2 —— 否则采集出的指纹可能与匹配器不兼容，规则库会被污染。 + +验收：用户框选正在播放的广告区间后，同一广告在下一集出现时能被指纹匹配到并按 `PROMPT` 模式提示。 + +### Phase 3：AI 兜底 + 输入升级 + +范围：`AiClassifier`、`AiAdDetectionService` 的 prompt 升级、`AiDebugLog` 的敏感信息收紧。 + +验收：本地通道全部弃权时 AI 被调用并给出可读解释；本地通道有强证据时 AI **不被**调用（省 token、省延迟）。 + +### 明确不做 + +- 服务端上报与众包聚合 +- 视觉 / OCR / 帧分析 +- 直播区间反馈 +- IJK / MPV 音频指纹 +- 自动启用高风险规则 + +## 14. 风险与缓解 + +| 风险 | 影响 | 缓解 | +|---|---|---| +| 用户框选不准，生成错误规则 | 正片被误删 | 规则强制带 `playlistHostSuffixes` + `siteKey` 收窄作用域；`minimumSignals ≥ 2`；`HlsManifestCleaner` 的既有安全阈值（删除比例 > 35%、时长 > 90s 一律回退）仍然兜底 | +| 归因结论全是 `SESSION_SKIP_ONLY` | 功能沦为手动跳过按钮 | 用 11 节的 `planAcceptedByKind` 度量；若一年后仍如此则回退简化 | +| 规则库膨胀 | 每 manifest 匹配耗时上升 | 碰撞检查拒绝重复；`HlsRuleConfig` 已有编译快照缓存；单条规则可禁用 | +| `legacyHeuristicActive` 干扰归因 | 证据链错误 | `ExistingRuleClassifier` 显式检测并在证据里注明；不改 AAR | +| 三端改造引入回归 | 播放器功能受损 | 源码断言测试锁定；Phase 1 带开关可整体回退 | +| 音频指纹与匹配器不兼容 | 采集的规则永不命中 | 黄金向量比对作为 Phase 2 硬前置 | + +## 15. 待确认决策 + +1. **短按的默认回溯窗口取 90 秒是否合适？** 太短会漏掉长贴片，太长会把正片框进来。建议做成设置项，默认 90s。 +2. **`AiDebugLog` 是否改为仅 debug 构建启用？** 这是既有问题，本设计只是让它暴露更多上下文。倾向于改，但属于扩大范围，需批准。 +3. **`legacyFallback=true` 是否要改为可配置？** 2.3 指出当前实际主力是无规则启发式。让用户能关掉它有利于归因准确性，但可能降低开箱去广效果。建议 Phase 1 只做诊断展示，不改行为。 +4. **`SitePlaylistHostBaseline` 的全局 200 站点上限是否足够？** 重度用户可能超。可调。 + +## 16. 架构决策摘要 + +### 决策 A：区间提交同时授权「本次执行」与「候选生成」，二者分离 + +用户框选区间这个动作本身就是对该片段的明确授权，当次跳过无需二次确认；但把它变成长期规则会影响未来所有播放，必须单独确认。因此第一屏的「本次立即跳过」默认勾选并立即执行，第二屏的「保存为规则」默认不选。 + +这样既解决了「反馈完什么都没得到」的体验问题，又不违反「AI/自动检测不得直接执行」的既有约束 —— 因为执行的授权来自用户的显式框选，不是来自任何检测器。 + +### 决策 B：AI 从裁判降级为兜底解释者 + +本地通道（切片结构、域名、已有规则）有确定性证据，大模型在这些场景只会引入噪声、延迟和 token 成本。只在本地全部弃权时才调 AI。副作用是按钮不再依赖 AI 配置，可见性门槛大幅降低，这本身是收益。 + +### 决策 C：仲裁按成本-风险排序，不按置信度裸排 + +「最好的去广告方式」不等于「置信度最高的方式」。启用一条已有的内置规则永远优于新建规则；域名黑名单永远优于音频指纹。排序权重把这个偏好编码进仲裁器。 + +### 决策 D：新逻辑全部下沉到 `main` 源集，三端只留 Host 适配 + +现有三份重复实现是维护负担的根源（改行为要同步改三处，测试靠字符串匹配锁定）。本次改造顺带解决，但不做无关重构 —— 只搬「有广告」这一条链路。 + +### 决策 E：不为区间反馈引入任何新的持久化后端 + +`SitePlaylistHostBaseline` 用 SharedPreferences，候选复用 `ImportedAdRuleCandidate` 的字段结构（它已有 `classification` / `confidence` / `riskLevel` / `reasons` / `status`），指纹复用 `AdAudioRuleStore` 的文件。全项目广告相关数据零 Room 表，这个性质保持不变。 From f19e81785a88899ea53351c28a368220f66f0f4b Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 05:13:43 +0800 Subject: [PATCH 18/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E7=AC=AC?= =?UTF-8?q?=E4=BA=8C=E8=BD=AE=E8=AF=84=E5=AE=A1=EF=BC=8C=E5=9F=9F=E5=90=8D?= =?UTF-8?q?=E9=80=9A=E9=81=93=E4=BC=9A=E7=94=9F=E6=88=90=E5=81=9C=E6=8E=89?= =?UTF-8?q?=E5=85=A8=E7=AB=99=E5=8E=BB=E5=B9=BF=E7=9A=84=E6=9C=89=E5=AE=B3?= =?UTF-8?q?=E8=A7=84=E5=88=99?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 复审发现上一轮声称修好的死代码仍是死的，另有一条会让情况变得更糟： 1. recordAdFeedbackHost() 写成 if (mAdFeedbackHost != null)，而适配器只在 用户首次按下「有广告」时懒创建 —— 首次反馈前基线永远为空，域名通道照旧 因「无基线不断言」弃权。改为强制走 adFeedback() 初始化。 2. 域名通道缺 HLS 通道那道对照校验。playlist 在 api.site.com、切片全在 cdn.site-x.com 是常见拆分，此时基线只有 playlist host，outsideBaseline 恒真 → 生成命中全站切片的规则 → HlsManifestCleaner 因 removedCount == segmentCount 回退原文，而 HlsAdblockPipeline 的 fallback 会连带停掉 legacy 启发式，比不加规则更差。新增 hasSameDomainOutside() 要求区间外存在同域切片作对照。 3. pathPatternsOf 只校验区间内每片都含目录段，未校验区间外不含。规则按切片 独立匹配，没有「只在区间内生效」的概念，站内正片路径恰好含 hint 时会被 一起删。改为双向校验。 4. classify 的守卫从 encodableCount() == 0 改为以实际载荷是否为空为准 —— 信号命中不等于条件有区分度。同时把 discontinuity 从 encodableCount 摘掉， 与 payloadOf 不编码它保持一致。 5. 换画质/换解码绕过 resetAdFeedback：这类「同集换 URL」不经 getPlayer 或 stopInlinePlayerForReload，切片结构与在途归因都属于上一画质。三端补接。 6. resetAdFeedback 补 close() 对话框：invalidate() 后归因回调不再刷新界面， 对话框会永久停在「分析中」。 7. durationMs() 补 try/catch 与 safePositionMs 对称；统计移到成交之后， 未成立的反馈不再计入。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback: 122 tests 0 failures; full suite 4018 tests with the same 5 pre-existing failures as baseline; compileMobileArm64_v8aDebug + compileLeanbackArm64_v8aDebug OK Task-Guard: AF5d-review2 --- .../android/tv/ui/activity/VideoActivity.java | 24 ++++++++++-- .../tv/ad/feedback/AdFeedbackHostAdapter.java | 8 +++- .../feedback/DomainReputationClassifier.java | 17 +++++++++ .../tv/ad/feedback/HlsSegmentClassifier.java | 38 ++++++++++++++----- .../tv/ui/activity/TmdbDetailActivity.java | 25 ++++++++++-- .../android/tv/ui/activity/VideoActivity.java | 26 +++++++++++-- .../DomainReputationClassifierTest.java | 23 +++++++++++ .../tv/ad/feedback/RulePayloadTest.java | 22 +++++++++++ 8 files changed, 160 insertions(+), 23 deletions(-) diff --git a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index b8d753bc35..b97544eeaa 100644 --- a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -2920,6 +2920,8 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; updateActionQuality(result); return; } + // 换画质同集换 URL：切片结构与在途归因都属于上一画质 + resetAdFeedback(); mQualityAdapter.setPosition(position); updateActionQuality(result); onItemClick(result); @@ -4026,17 +4028,26 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; mAdMarkStartMs = -1; if (mAdFeedback != null) mAdFeedback.invalidate(); if (mAdFeedbackHost != null) mAdFeedbackHost.invalidateEvidence(); + // 作废后归因回调不再刷新界面，对话框会永久停在「分析中」 + if (mAdFeedbackDialog != null) mAdFeedbackDialog.close(); + mAdFeedbackDialog = null; } - /** 起播后记录本站域名，供域名信誉通道建立基线。 */ + /** + * 起播后记录本站域名，供域名信誉通道建立基线。 + * + *

必须强制初始化适配器：它原本只在用户首次按下「有广告」时懒创建， + * 若这里跳过未初始化的情况，首次反馈时基线仍为空，域名通道会因 + * 「无基线不断言」而弃权 —— 整条通道等于没接。 + */ private void recordAdFeedbackHost() { - if (mAdFeedbackHost != null) mAdFeedbackHost.recordPlaybackHost(); + adFeedback(); + mAdFeedbackHost.recordPlaybackHost(); } /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitAdFeedback() { AdFeedbackController controller = adFeedback(); - AdBlockStatsStore.recordFeedback(getKey()); AdFeedbackSession session; if (mAdMarkStartMs >= 0) { // 先取位置再清标记：播放器已释放时保留标记，用户可重试而不必重新打点 @@ -4051,7 +4062,12 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; } else { session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } - if (session == null) Notify.show(R.string.ad_interval_invalid); + if (session == null) { + Notify.show(R.string.ad_interval_invalid); + return; + } + // 统计放在成交之后，未成立的反馈不计入 + AdBlockStatsStore.recordFeedback(getKey()); } /** 长按：打起点进入标记模式，再次长按取消。 */ diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java index abb4c509b9..90e2c6d0cf 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java @@ -144,7 +144,13 @@ public final class AdFeedbackHostAdapter implements AdFeedbackController.Host { @Override public long durationMs() { - return playback.durationMs(); + // 与 safePositionMs 对称：PlayerManager.getDuration() 也是裸 + // player.getDuration()，播放器 release 后 player 为 null 会 NPE + try { + return playback.durationMs(); + } catch (RuntimeException e) { + return 0L; + } } @Override diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java index 6db88411b6..ddfef18bff 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java @@ -56,6 +56,13 @@ public final class DomainReputationClassifier { Set foreignHosts = foreignHosts(evidence); if (foreignHosts.isEmpty()) return null; + // 必须有对照组：区间外存在属于 playlist 域名的切片，才能说明这些外域 + // 切片是异常的。否则整站切片都在独立 CDN（api.site.com + cdn.site-x.com + // 这种常见拆分）会被整体判为广告，生成的规则命中全部切片 → + // HlsManifestCleaner 因 removedCount == segmentCount 回退原文， + // 而 HlsAdblockPipeline 的 fallback 会连带停掉 legacy 启发式， + // 结果比不加规则更差。与 HlsSegmentClassifier.detectCrossDomain 同一判据。 + if (!hasSameDomainOutside(evidence)) return null; List matchedBlacklist = matching(foreignHosts, safe.blacklistedHosts()); List matchedCandidates = matching(foreignHosts, safe.interfaceCandidateHosts()); @@ -106,6 +113,16 @@ public final class DomainReputationClassifier { confidence, RiskLevel.LOW, evidenceLines, remediation, payload); } + /** + * 区间外是否存在属于 playlist 域名的切片。这是「外域切片异常」的对照前提： + * 没有同域切片作对照，说明该站本来就把切片放在独立 CDN 上。 + */ + private static boolean hasSameDomainOutside(AdIntervalEvidence evidence) { + String playlistHost = evidence.playlistHost(); + if (playlistHost.isEmpty()) return false; + return evidence.outside().stream().anyMatch(fact -> fact.hostEndsWith(playlistHost)); + } + /** 区间内不属于 playlist 域名的切片 host，去重保序。 */ private static Set foreignHosts(AdIntervalEvidence evidence) { Set hosts = new LinkedHashSet<>(); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index 07efbade05..b78b8f5a7a 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -45,11 +45,17 @@ public final class HlsSegmentClassifier { public record Signals(boolean crossDomain, boolean discontinuity, boolean durationOutlier, boolean pathHint, boolean headPosition) { - /** 可编码进 {@code HlsAdRule} 的信号数。位置信号不被 HlsAdRule 支持，不计入。 */ + /** + * 可编码进 {@code HlsAdRule} 的信号数。 + * + *

不计入位置信号（HlsAdRule 不支持）与断点信号：断点只对广告块首片 + * 成立，编码进规则会导致只删首片，因此 payloadOf 刻意不编码它。 + * 这里必须与 payloadOf 保持一致，否则 classify 的守卫会放过只有 + * 断点+位置的场景，产出一条空载荷的归因。 + */ public int encodableCount() { int count = 0; if (crossDomain) count++; - if (discontinuity) count++; if (durationOutlier) count++; if (pathHint) count++; return count; @@ -77,12 +83,15 @@ public final class HlsSegmentClassifier { Signals signals = detect(evidence); float confidence = signals.weightSum(); if (confidence < MIN_CONFIDENCE) return null; - // 无法编码进规则的结论没有落地价值 - if (signals.encodableCount() == 0) return null; + + // 以实际载荷为准判断能否落地：信号命中不等于条件有区分度 + // （如站内正片路径也含 /ads/ 时，路径特征无法编码进规则） + RulePayload payload = payloadOf(evidence, signals); + if (payload.isEmpty()) return null; return new AdAttribution(CHANNEL_ID, categoryOf(signals), confidence, RiskLevel.MEDIUM, describe(evidence, signals), RemediationKind.HLS_STRUCTURED_RULE, - payloadOf(evidence, signals)); + payload); } /** @@ -106,7 +115,7 @@ public final class HlsSegmentClassifier { .distinct() .toList(); List pathPatterns = signals.pathHint() - ? pathPatternsOf(evidence.inside()) : List.of(); + ? pathPatternsOf(evidence.inside(), evidence.outside()) : List.of(); // 没有任何区分条件时不得生成规则 if (hosts.isEmpty() && pathPatterns.isEmpty()) return RulePayload.empty(); @@ -143,13 +152,22 @@ public final class HlsSegmentClassifier { false, requireCrossDomain, minimumSignals); } - /** 命中的广告路径目录段转成正则，用 quote 避免元字符注入。 */ - private static List pathPatternsOf(List inside) { + /** + * 命中的广告路径目录段转成正则，用 quote 避免元字符注入。 + * + *

双向校验：区间内每片都含该目录段，且区间外**没有**任何切片含它。 + * 只查 inside 的话，站内正片路径恰好含 hint 时会被一起删掉 —— 规则是按 + * 切片独立匹配的，没有「只在这段区间内生效」的概念。 + */ + private static List pathPatternsOf(List inside, List outside) { List patterns = new ArrayList<>(); for (String hint : PATH_HINTS) { - boolean allMatch = inside.stream() + boolean insideAll = inside.stream() .allMatch(fact -> fact.path().toLowerCase(Locale.US).contains(hint)); - if (allMatch) patterns.add(Pattern.quote(hint)); + if (!insideAll) continue; + boolean outsideNone = outside.stream() + .noneMatch(fact -> fact.path().toLowerCase(Locale.US).contains(hint)); + if (outsideNone) patterns.add(Pattern.quote(hint)); } return patterns; } diff --git a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java index 62e8a08a21..64bfccf6cd 100644 --- a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java +++ b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java @@ -6934,6 +6934,9 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. } currentInlineResult = result; useParse = result.shouldUseParse(); + // 换画质走 changeInlineQuality → startInlinePlayer，不经 + // stopInlinePlayerForReload，此处兜住同集换 URL 的作废 + resetAdFeedback(); if (result.hasPosition() && history != null) history.setPosition(result.getPosition()); if (result.hasDesc() && !hasTmdbOverview()) { vod.setContent(result.getDesc()); @@ -8326,17 +8329,26 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. mAdMarkStartMs = -1; if (mAdFeedback != null) mAdFeedback.invalidate(); if (mAdFeedbackHost != null) mAdFeedbackHost.invalidateEvidence(); + // 作废后归因回调不再刷新界面，对话框会永久停在「分析中」 + if (mAdFeedbackDialog != null) mAdFeedbackDialog.close(); + mAdFeedbackDialog = null; } - /** 起播后记录本站域名，供域名信誉通道建立基线。 */ + /** + * 起播后记录本站域名，供域名信誉通道建立基线。 + * + *

必须强制初始化适配器：它原本只在用户首次按下「有广告」时懒创建， + * 若这里跳过未初始化的情况，首次反馈时基线仍为空，域名通道会因 + * 「无基线不断言」而弃权 —— 整条通道等于没接。 + */ private void recordAdFeedbackHost() { - if (mAdFeedbackHost != null) mAdFeedbackHost.recordPlaybackHost(); + adFeedback(); + mAdFeedbackHost.recordPlaybackHost(); } /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitInlineAdFeedback() { AdFeedbackController controller = adFeedback(); - AdBlockStatsStore.recordFeedback(getKeyText()); AdFeedbackSession session; if (mAdMarkStartMs >= 0) { // 先取位置再清标记：播放器已释放时保留标记，用户可重试而不必重新打点 @@ -8351,7 +8363,12 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. } else { session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } - if (session == null) Notify.show(R.string.ad_interval_invalid); + if (session == null) { + Notify.show(R.string.ad_interval_invalid); + return; + } + // 统计放在成交之后，未成立的反馈不计入 + AdBlockStatsStore.recordFeedback(getKeyText()); } /** 长按：打起点进入标记模式，再次长按取消。 */ diff --git a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index 1931b78dd9..dd295f97a8 100644 --- a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -2759,6 +2759,8 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco updateActionQuality(result); return; } + // 换画质同集换 URL：切片结构与在途归因都属于上一画质 + resetAdFeedback(); mQualityAdapter.setPosition(position); updateActionQuality(result); onItemClick(result); @@ -4301,6 +4303,8 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco private void switchDecodeWithResult(Result result, long position, float speed, boolean repeat, MediaMetadata metadata) { decodeSwitchRefreshing = false; + // 换解码可能重新解析出不同 URL，与换画质同理 + resetAdFeedback(); if (result == null || result.hasMsg() || result.getRealUrl().isEmpty()) { player().toggleDecode(); } else { @@ -8757,17 +8761,26 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco mAdMarkStartMs = -1; if (mAdFeedback != null) mAdFeedback.invalidate(); if (mAdFeedbackHost != null) mAdFeedbackHost.invalidateEvidence(); + // 作废后归因回调不再刷新界面，对话框会永久停在「分析中」 + if (mAdFeedbackDialog != null) mAdFeedbackDialog.close(); + mAdFeedbackDialog = null; } - /** 起播后记录本站域名，供域名信誉通道建立基线。 */ + /** + * 起播后记录本站域名，供域名信誉通道建立基线。 + * + *

必须强制初始化适配器：它原本只在用户首次按下「有广告」时懒创建， + * 若这里跳过未初始化的情况，首次反馈时基线仍为空，域名通道会因 + * 「无基线不断言」而弃权 —— 整条通道等于没接。 + */ private void recordAdFeedbackHost() { - if (mAdFeedbackHost != null) mAdFeedbackHost.recordPlaybackHost(); + adFeedback(); + mAdFeedbackHost.recordPlaybackHost(); } /** 短按：标记模式下第二次按下确定终点，否则以当前位置为终点回溯推断起点。 */ private void submitAdFeedback() { AdFeedbackController controller = adFeedback(); - AdBlockStatsStore.recordFeedback(getKey()); AdFeedbackSession session; if (mAdMarkStartMs >= 0) { // 先取位置再清标记：播放器已释放时保留标记，用户可重试而不必重新打点 @@ -8782,7 +8795,12 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco } else { session = controller.onQuickReport(mAdFeedbackHost.cachedEvidence()); } - if (session == null) Notify.show(R.string.ad_interval_invalid); + if (session == null) { + Notify.show(R.string.ad_interval_invalid); + return; + } + // 统计放在成交之后，未成立的反馈不计入 + AdBlockStatsStore.recordFeedback(getKey()); } /** 长按：打起点进入标记模式，再次长按取消。 */ diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java index faf4f2e45f..fbd81e09d1 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java @@ -83,6 +83,29 @@ public class DomainReputationClassifierTest { assertTrue(attribution.evidence().stream().anyMatch(line -> line.contains("饭太硬"))); } + @Test + public void abstainsWhenSiteHasNoSameDomainSegmentsAtAll() { + // playlist 在 v.example.com、全部切片在独立 CDN，是常见的架构拆分。 + // 没有同域切片作对照就断言「外域即广告」，生成的规则会命中全站切片 → + // cleaner 因 removedCount == segmentCount 回退原文，而 + // HlsAdblockPipeline 的 fallback 会连带停掉 legacy 启发式，比不加规则更差。 + List inside = List.of( + new SegmentFact(3, "cdn.site-x.com", "/seg/3.ts", 6.4, false)); + List outside = List.of( + new SegmentFact(0, "cdn.site-x.com", "/seg/0.ts", 8.0, false), + new SegmentFact(9, "cdn.site-x.com", "/seg/9.ts", 8.0, false)); + AdIntervalEvidence evidence = new AdIntervalEvidence( + 10_000, 40_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + PLAYLIST_HOST, "/play/index.m3u8", true, + inside, outside, false, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + assertNull(DomainReputationClassifier.classify(evidence, + new DomainReputationClassifier.Input( + List.of(), List.of(PLAYLIST_HOST), List.of(), ""))); + } + @Test public void abstainsWithoutBaselineToAvoidFirstPlayMisjudgement() { // 首次播放该站，没有任何基线数据：不能断言域名异常 diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java index 741efc6a7c..f333d93472 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java @@ -2,6 +2,7 @@ package com.fongmi.android.tv.ad.feedback; import static org.junit.Assert.assertEquals; import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertNull; import static org.junit.Assert.assertTrue; import org.junit.Test; @@ -119,6 +120,27 @@ public class RulePayloadTest { assertFalse(payload.hasDurationRange()); } + @Test + public void pathHintPresentOutsideTheIntervalIsNotADistinguishingCondition() { + // 站内正片路径也含 /ads/ 时，该特征无区分度：规则按切片独立匹配， + // 没有「只在这段区间内生效」的概念，编码它会连正片一起删。 + List inside = List.of( + new SegmentFact(10, "v.example.com", "/ads/10.ts", 8.0, true), + new SegmentFact(11, "v.example.com", "/ads/11.ts", 8.0, false)); + List outside = List.of( + new SegmentFact(0, "v.example.com", "/ads/0.ts", 8.0, false), + new SegmentFact(12, "v.example.com", "/seg/12.ts", 8.0, false)); + AdIntervalEvidence evidence = new AdIntervalEvidence( + 80_000, 96_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + "v.example.com", "/play/index.m3u8", true, + inside, outside, true, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + // 同域 + 时长一致 + 路径无区分度 → 无任何可编码条件，必须弃权 + assertNull(HlsSegmentClassifier.classify(evidence)); + } + @Test public void existingRuleAttributionCarriesRuleKey() { ExistingRuleClassifier.RuleState disabled = new ExistingRuleClassifier.RuleState( From d79d7d35765580290cca2c14bae13262d0f4d7ad Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 05:21:01 +0800 Subject: [PATCH 19/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9ATmdbDetail=20?= =?UTF-8?q?=E6=96=AD=E8=BF=9E=E6=81=A2=E5=A4=8D=E4=B8=8D=E5=BA=94=E5=85=B3?= =?UTF-8?q?=E6=8E=89=E7=94=A8=E6=88=B7=E6=AD=A3=E5=9C=A8=E7=9C=8B=E7=9A=84?= =?UTF-8?q?=E5=8F=8D=E9=A6=88=E5=AF=B9=E8=AF=9D=E6=A1=86?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 上一个提交把 resetAdFeedback() 放进 startInlinePlayer，但该方法同时承担 断连恢复（restoreInlinePlaybackIfNeeded / startPendingInlinePlayer）与 正常换集（playInline → onParseSuccess → startInlinePlayer）。后果是： - 正常换集会重复作废两次（playInline 已经过 stopInlinePlayerForReload）； - 服务重连时会关掉用户此刻正开着的归因对话框。 改为只在 changeInlineQuality 里作废 —— 那才是「同集换 URL」的唯一入口。 换集路径已由 stopInlinePlayerForReload 覆盖，无需在汇聚点重复。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback + ui.activity: 0 failures; compileMobileArm64_v8aDebug OK Task-Guard: AF5e-reset-guard --- .../fongmi/android/tv/ui/activity/TmdbDetailActivity.java | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java index 64bfccf6cd..b69fb05655 100644 --- a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java +++ b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java @@ -6934,9 +6934,6 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. } currentInlineResult = result; useParse = result.shouldUseParse(); - // 换画质走 changeInlineQuality → startInlinePlayer，不经 - // stopInlinePlayerForReload，此处兜住同集换 URL 的作废 - resetAdFeedback(); if (result.hasPosition() && history != null) history.setPosition(result.getPosition()); if (result.hasDesc() && !hasTmdbOverview()) { vod.setContent(result.getDesc()); @@ -7789,6 +7786,10 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. return; } saveInlineHistory(); + // 换画质同集换 URL：切片结构与在途归因都属于上一画质。 + // 放在这里而不是 startInlinePlayer，因为后者也承担断连恢复， + // 那条路径不该关掉用户正在看的反馈对话框。 + resetAdFeedback(); currentInlineResult.getUrl().set(position); updateInlineButtons(service() != null && player() != null && !player().isEmpty() && player().isPlaying()); startInlinePlayer(currentInlineResult); From b3ea90384bcd20fd0eadd9a1c3ef1b186a60f575 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 06:55:55 +0800 Subject: [PATCH 20/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E7=AC=AC?= =?UTF-8?q?=E4=B8=89=E8=BD=AE=E8=AF=84=E5=AE=A1=EF=BC=8CHLS=20=E9=80=9A?= =?UTF-8?q?=E9=81=93=E4=BB=8D=E4=BC=9A=E7=94=9F=E6=88=90=E5=81=9C=E6=8E=89?= =?UTF-8?q?=E5=85=A8=E7=AB=99=E5=8E=BB=E5=B9=BF=E7=9A=84=E8=A7=84=E5=88=99?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 上一轮只给域名通道补了对照校验，HLS 通道的同一漏洞没堵住 —— 评审实测 「playlist 在 api.site.com、切片全在 cdn.site-x.com、区间首片带断点且落在 片头」这个常见形态，置信度 0.30 恰好压线过关，产出 minimumSignals=1 + 全站 CDN 域名的规则，命中每一个切片。 1. payloadOf 的 hosts 现在只在 signals.crossDomain() 成立时编码。crossDomain 自带对照校验（要求区间外存在同域切片），少了这道门槛就会把全站域名写进 规则。此前只筛「不是 playlist 域名」，与 pathPatterns 的双向校验不对称。 2. pathPatternsOf 在区间外为空时弃权：用户框选覆盖整个 playlist 时无法完成 对照，任何条件都命中全部切片。 3. hasSameDomainOutside 守卫下移到黑名单分支之后。黑名单分支返回的是 NONE + 空载荷的纯诊断，不可能生成规则，不该被这道守卫拦掉 —— 而「整站 切片在独立 CDN」恰是那条诊断最该出现的场景。 4. describe() 新增 describeRule()，逐项列出实际写进规则的条件。此前证据按 signals 输出，与实际编码脱钩：用户看到「路径包含广告目录特征」却不知道 规则其实按 host 生效，也看不到规则的杀伤范围。 5. 三端换播放器内核补 resetAdFeedback()：它重新解析出可能不同的 URL， 与换画质同理。此前只接了换画质与换解码。 6. recordAdFeedbackHost() 加开关判断，关闭去广告或播直播时不建基线， 避免这些域名挤掉 LRU 里有用的条目。 7. SitePlaylistHostBaseline.record 在 host 已是最近一条时跳过写盘 —— STATE_READY 会因 seek、卡顿恢复反复触发。 8. Signals.encodableCount() 更名 signalCount() 并恢复计入全部信号：它现在 只用于诊断展示，不再参与落地判断，避免下次又出现与 payloadOf 的语义漂移。 GenRuleE2ETest 补两条回归：全站 CDN 与区间覆盖整个 playlist，两个通道都 必须弃权或产出不可落地的结论。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback: 124 tests 0 failures (GenRuleE2ETest 5, +2 site-wide-CDN regressions); full suite 4020 tests with the same 5 pre-existing failures as baseline; compileMobileArm64_v8aDebug + compileLeanbackArm64_v8aDebug OK Task-Guard: AF5f-review3 --- .../android/tv/ui/activity/VideoActivity.java | 4 ++ .../feedback/DomainReputationClassifier.java | 36 +++++----- .../tv/ad/feedback/HlsSegmentClassifier.java | 72 ++++++++++++++----- .../ad/feedback/SitePlaylistHostBaseline.java | 9 ++- .../tv/ui/activity/TmdbDetailActivity.java | 4 ++ .../android/tv/ui/activity/VideoActivity.java | 4 ++ .../tv/ad/feedback/GenRuleE2ETest.java | 46 ++++++++++++ .../ad/feedback/HlsSegmentClassifierTest.java | 13 ++-- 8 files changed, 149 insertions(+), 39 deletions(-) diff --git a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index b97544eeaa..a89b51a8e9 100644 --- a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -3954,6 +3954,8 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; private void switchPlayerKernelWithResult(int type, Result result, long position, float speed, boolean repeat, MediaMetadata metadata) { playerKernelSwitchRefreshing = false; + // 换内核会重新解析出可能不同的 URL，与换画质同理 + resetAdFeedback(); if (result == null || result.hasMsg() || result.getRealUrl().isEmpty()) { Notify.show(result != null && result.hasMsg() ? result.getMsg() : getString(R.string.error_play_url)); } else { @@ -4041,6 +4043,8 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; * 「无基线不断言」而弃权 —— 整条通道等于没接。 */ private void recordAdFeedbackHost() { + // 关掉去广告或播直播时不建基线：这些域名会挤掉 LRU 里真正有用的条目 + if (!isAdFeedbackEnabled()) return; adFeedback(); mAdFeedbackHost.recordPlaybackHost(); } diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java index ddfef18bff..22abed8e8f 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java @@ -56,20 +56,15 @@ public final class DomainReputationClassifier { Set foreignHosts = foreignHosts(evidence); if (foreignHosts.isEmpty()) return null; - // 必须有对照组：区间外存在属于 playlist 域名的切片，才能说明这些外域 - // 切片是异常的。否则整站切片都在独立 CDN（api.site.com + cdn.site-x.com - // 这种常见拆分）会被整体判为广告，生成的规则命中全部切片 → - // HlsManifestCleaner 因 removedCount == segmentCount 回退原文， - // 而 HlsAdblockPipeline 的 fallback 会连带停掉 legacy 启发式， - // 结果比不加规则更差。与 HlsSegmentClassifier.detectCrossDomain 同一判据。 - if (!hasSameDomainOutside(evidence)) return null; List matchedBlacklist = matching(foreignHosts, safe.blacklistedHosts()); List matchedCandidates = matching(foreignHosts, safe.interfaceCandidateHosts()); if (!matchedBlacklist.isEmpty()) { // 已在黑名单里却仍被用户看到：拦截路径没覆盖播放器直连的切片请求。 - // 结论是诊断，落地方式应改用 HLS 结构化规则。 + // 结论是诊断（NONE + 空载荷，actionable() 为 false），不会生成规则， + // 因此不受下面那道「需要同域对照」的约束 —— 整站切片都在独立 CDN + // 恰恰是这条诊断最该出现的场景。 List evidenceLines = new ArrayList<>(); evidenceLines.add("切片域名已在广告黑名单中：" + String.join("、", matchedBlacklist)); evidenceLines.add("黑名单主要在 WebView 请求拦截生效，不拦播放器直连的切片请求"); @@ -78,6 +73,15 @@ public final class DomainReputationClassifier { CONFIDENCE_KNOWN_HOST, RiskLevel.LOW, evidenceLines, RemediationKind.NONE); } + // 以下分支会生成规则，必须有对照组：区间外存在属于 playlist 域名的切片， + // 才能说明这些外域切片是异常的。否则整站切片都在独立 CDN + // （api.site.com + cdn.site-x.com 这种常见拆分）会被整体判为广告， + // 生成的规则命中全部切片 → HlsManifestCleaner 因 + // removedCount == segmentCount 回退原文，而 HlsAdblockPipeline 的 + // fallback 会连带停掉 legacy 启发式，比不加规则更差。 + // 与 HlsSegmentClassifier.detectCrossDomain 同一判据。 + if (!hasSameDomainOutside(evidence)) return null; + // 无基线数据时不能断言「非本站域名」，避免首次播放即误判 if (safe.siteBaselineHosts().isEmpty() && matchedCandidates.isEmpty()) return null; @@ -101,16 +105,14 @@ public final class DomainReputationClassifier { // 请求拦截生效（RuleConfig.getAds() 的唯一消费者是 CustomWebView）， // 拦不住播放器直连的切片请求。跨域按 foreignHosts 的定义天然成立， // 与 hostSuffixes 合起来正好满足 minimumSignals=2。 - List scope = evidence.playlistHost().isEmpty() - ? List.of() : List.of(evidence.playlistHost()); - RemediationKind remediation = scope.isEmpty() - ? RemediationKind.HOST_BLACKLIST : RemediationKind.HLS_STRUCTURED_RULE; - RulePayload payload = scope.isEmpty() - ? RulePayload.ofHosts(List.copyOf(foreignHosts)) - : RulePayload.ofHlsRule(scope, List.copyOf(foreignHosts), - Double.NaN, Double.NaN, false, true, 2); + // playlistHost 必然非空 —— hasSameDomainOutside 在它为空时已经弃权， + // 所以这里不再有「无作用域」的退路。 + List scope = List.of(evidence.playlistHost()); + RulePayload payload = RulePayload.ofHlsRule(scope, List.copyOf(foreignHosts), + Double.NaN, Double.NaN, false, true, 2); return new AdAttribution(CHANNEL_ID, AdCategory.THIRD_PARTY_CDN_SEGMENT, - confidence, RiskLevel.LOW, evidenceLines, remediation, payload); + confidence, RiskLevel.LOW, evidenceLines, + RemediationKind.HLS_STRUCTURED_RULE, payload); } /** diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index b78b8f5a7a..c1987d6965 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -46,18 +46,19 @@ public final class HlsSegmentClassifier { boolean pathHint, boolean headPosition) { /** - * 可编码进 {@code HlsAdRule} 的信号数。 + * 观察到的信号数，仅供诊断展示。 * - *

不计入位置信号（HlsAdRule 不支持）与断点信号：断点只对广告块首片 - * 成立，编码进规则会导致只删首片，因此 payloadOf 刻意不编码它。 - * 这里必须与 payloadOf 保持一致，否则 classify 的守卫会放过只有 - * 断点+位置的场景，产出一条空载荷的归因。 + *

刻意不用于判断能否落地 —— 信号命中不等于条件有区分度， + * 那件事由 {@code payloadOf} 的实际载荷决定。此前用它做守卫时， + * 「同域 + 路径特征区间外也有」的场景会产出一条空载荷的归因。 */ - public int encodableCount() { + public int signalCount() { int count = 0; if (crossDomain) count++; + if (discontinuity) count++; if (durationOutlier) count++; if (pathHint) count++; + if (headPosition) count++; return count; } @@ -90,8 +91,8 @@ public final class HlsSegmentClassifier { if (payload.isEmpty()) return null; return new AdAttribution(CHANNEL_ID, categoryOf(signals), confidence, - RiskLevel.MEDIUM, describe(evidence, signals), RemediationKind.HLS_STRUCTURED_RULE, - payload); + RiskLevel.MEDIUM, describe(evidence, signals, payload), + RemediationKind.HLS_STRUCTURED_RULE, payload); } /** @@ -107,13 +108,19 @@ public final class HlsSegmentClassifier { */ private static RulePayload payloadOf(AdIntervalEvidence evidence, Signals signals) { String playlistHost = evidence.playlistHost(); - // 只保留非本站域名 —— 同域 host 不是区分条件 - List hosts = evidence.inside().stream() - .map(SegmentFact::host) - .filter(host -> !host.isEmpty()) - .filter(host -> playlistHost.isEmpty() || !hostMatches(host, playlistHost)) - .distinct() - .toList(); + // 域名只在跨域信号成立时才有区分度。crossDomain 自带对照校验 + // （detectCrossDomain 要求区间外存在同域切片），少了这道门槛就会在 + // 「整站切片都在独立 CDN」的架构上把全站域名写进规则 —— 规则命中所有 + // 切片 → cleaner 因 removedCount == segmentCount 回退 → 连带停掉 + // legacy 启发式，比不加规则更差。 + List hosts = signals.crossDomain() + ? evidence.inside().stream() + .map(SegmentFact::host) + .filter(host -> !host.isEmpty()) + .filter(host -> playlistHost.isEmpty() || !hostMatches(host, playlistHost)) + .distinct() + .toList() + : List.of(); List pathPatterns = signals.pathHint() ? pathPatternsOf(evidence.inside(), evidence.outside()) : List.of(); // 没有任何区分条件时不得生成规则 @@ -158,8 +165,12 @@ public final class HlsSegmentClassifier { *

双向校验：区间内每片都含该目录段，且区间外**没有**任何切片含它。 * 只查 inside 的话，站内正片路径恰好含 hint 时会被一起删掉 —— 规则是按 * 切片独立匹配的，没有「只在这段区间内生效」的概念。 + * + *

区间外为空时无法完成对照（用户框选覆盖了整个 playlist），一律弃权： + * 此时任何条件都命中全部切片，生成的规则会触发 cleaner 整体回退。 */ private static List pathPatternsOf(List inside, List outside) { + if (outside.isEmpty()) return List.of(); List patterns = new ArrayList<>(); for (String hint : PATH_HINTS) { boolean insideAll = inside.stream() @@ -241,7 +252,15 @@ public final class HlsSegmentClassifier { return AdCategory.UNKNOWN; } - private static List describe(AdIntervalEvidence evidence, Signals signals) { + /** + * 生成给用户看的证据。 + * + *

分两段：先说观察到的现象（按 signals），再明确列出**实际写进规则**的 + * 条件（按 payload）。两者不是一回事 —— 例如路径特征因区间外也含而未被编码时， + * 现象成立但规则里没有它。不列清规则条件，用户无法判断这条规则的杀伤范围。 + */ + private static List describe(AdIntervalEvidence evidence, Signals signals, + RulePayload payload) { List inside = evidence.inside(); List lines = new ArrayList<>(); lines.add(String.format(Locale.US, "区间内 %d 个切片，时长合计 %.1fs", @@ -258,6 +277,27 @@ public final class HlsSegmentClassifier { if (signals.pathHint()) lines.add("切片路径包含广告目录特征"); if (signals.headPosition()) lines.add("区间位于播放列表头部"); lines.add("起点来源：" + evidence.startOrigin()); + lines.addAll(describeRule(payload)); + return lines; + } + + /** 逐项列出规则实际生效的条件，让用户能判断杀伤范围。 */ + private static List describeRule(RulePayload payload) { + List lines = new ArrayList<>(); + if (payload.isEmpty()) return lines; + lines.add("规则生效范围：" + String.join("、", payload.playlistHostSuffixes())); + if (!payload.hosts().isEmpty()) { + lines.add("删除来自这些域名的切片：" + String.join("、", payload.hosts())); + } + if (!payload.regex().isEmpty()) { + lines.add("删除路径匹配的切片：" + String.join("、", payload.regex())); + } + if (payload.hasDurationRange()) { + lines.add(String.format(Locale.US, "限定切片时长 %.2f–%.2fs", + payload.durationMin(), payload.durationMax())); + } + if (payload.requireCrossDomain()) lines.add("仅当切片域名与 playlist 不同时生效"); + lines.add(String.format(Locale.US, "需同时满足其中 %d 个条件", payload.minimumSignals())); return lines; } diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java index 21fd6ccabf..3621795114 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java @@ -59,20 +59,27 @@ public final class SitePlaylistHostBaseline { public synchronized void record(String siteKey, String... hosts) { if (siteKey == null || siteKey.isBlank() || hosts == null) return; List known = sites.computeIfAbsent(siteKey, key -> new ArrayList<>()); + boolean seenAny = false; boolean changed = false; for (String host : hosts) { if (host == null || host.isBlank()) continue; String normalized = host.toLowerCase(Locale.US).trim(); + seenAny = true; + // 已在末尾说明状态没变，跳过写盘 —— STATE_READY 会因 seek、 + // 卡顿恢复反复触发，每次都 persist 是不必要的 I/O + int last = known.size() - 1; + if (last >= 0 && known.get(last).equals(normalized)) continue; // 已存在则移到末尾，表示最近使用 known.remove(normalized); known.add(normalized); changed = true; } - if (!changed) { + if (!seenAny) { // 全部输入为空白：不能留下空条目占用站点配额 if (known.isEmpty()) sites.remove(siteKey); return; } + if (!changed) return; while (known.size() > MAX_HOSTS_PER_SITE) known.remove(0); evictSites(); persist(); diff --git a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java index b69fb05655..d4003b8551 100644 --- a/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java +++ b/app/src/main/java/com/fongmi/android/tv/ui/activity/TmdbDetailActivity.java @@ -8234,6 +8234,8 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. finishInlinePlayerSwitchRequest(); Notify.show(result != null && result.hasMsg() ? result.getMsg() : getString(R.string.error_play_url)); } else { + // 换内核重新解析出可能不同的 URL，与换画质同理 + resetAdFeedback(); currentInlineResult = result; inlineHttpRefreshAttempted = false; useParse = result.shouldUseParse(); @@ -8343,6 +8345,8 @@ public class TmdbDetailActivity extends PlaybackActivity implements TrackDialog. * 「无基线不断言」而弃权 —— 整条通道等于没接。 */ private void recordAdFeedbackHost() { + // 关掉去广告或播直播时不建基线：这些域名会挤掉 LRU 里真正有用的条目 + if (!isInlineAdFeedbackEnabled()) return; adFeedback(); mAdFeedbackHost.recordPlaybackHost(); } diff --git a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index dd295f97a8..47ff94d213 100644 --- a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -4438,6 +4438,8 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco private void switchPlayerKernelWithResult(int requestId, int type, Result result, long position, float speed, boolean repeat, MediaMetadata metadata) { if (requestId != playerKernelSwitchRequestId) return; + // 换内核会重新解析出可能不同的 URL，与换画质/换解码同理 + resetAdFeedback(); if (result == null || result.hasMsg() || result.getRealUrl().isEmpty()) { Notify.show(result != null && result.hasMsg() ? result.getMsg() : getString(R.string.error_play_url)); } else { @@ -8774,6 +8776,8 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco * 「无基线不断言」而弃权 —— 整条通道等于没接。 */ private void recordAdFeedbackHost() { + // 关掉去广告或播直播时不建基线：这些域名会挤掉 LRU 里真正有用的条目 + if (!isAdFeedbackEnabled()) return; adFeedback(); mAdFeedbackHost.recordPlaybackHost(); } diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/GenRuleE2ETest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/GenRuleE2ETest.java index fe74c98b81..e60840e169 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/GenRuleE2ETest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/GenRuleE2ETest.java @@ -1,7 +1,9 @@ package com.fongmi.android.tv.ad.feedback; import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; import static org.junit.Assert.assertNotNull; +import static org.junit.Assert.assertNull; import com.fongmi.android.tv.bean.HlsAdRule; import com.fongmi.android.tv.utils.HlsManifestCleaner; @@ -117,4 +119,48 @@ public class GenRuleE2ETest { assertEquals("只能删广告块，不得误删断点后的正片", AD_COUNT, res.removedSegments()); } + + /** + * 整站切片都在独立 CDN（api.site.com + cdn.site-x.com 这种常见拆分）。 + * 没有同域切片作对照时，任何域名条件都命中全站 → cleaner 因 + * removedCount == segmentCount 回退，而 HlsAdblockPipeline 的 fallback + * 会连带停掉 legacy 启发式，比不加规则更差。两个通道都必须弃权。 + */ + @Test + public void siteWideCdnMustNotProduceAnyRule() { + List inside = seg(AD_FROM, AD_COUNT, "cdn.site-x.com", "/seg/", 6.4, true); + List outside = new ArrayList<>( + seg(0, AD_FROM, "cdn.site-x.com", "/seg/", 8.0, false)); + outside.addAll(seg(AD_FROM + AD_COUNT, TOTAL - AD_FROM - AD_COUNT, + "cdn.site-x.com", "/seg/", 8.0, true)); + AdIntervalEvidence evidence = new AdIntervalEvidence( + 80_000, 100_000, StartOrigin.USER_MARKED, + "site", "站点", "剧", "线", "1", "api.site.com", "/play/index.m3u8", true, + inside, outside, true, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + AdAttribution hls = HlsSegmentClassifier.classify(evidence); + if (hls != null) { + assertFalse("HLS 通道不得为全站 CDN 生成可落地规则", hls.actionable()); + } + assertNull("域名通道必须弃权", DomainReputationClassifier.classify(evidence, + new DomainReputationClassifier.Input( + List.of(), List.of("api.site.com"), List.of(), ""))); + } + + /** 用户框选覆盖整个 playlist 时无对照可用，同样不得生成规则。 */ + @Test + public void intervalCoveringWholePlaylistMustNotProduceAnyRule() { + AdIntervalEvidence evidence = new AdIntervalEvidence( + 0, 240_000, StartOrigin.USER_MARKED, + "site", "站点", "剧", "线", "1", "v.example.com", "/play/index.m3u8", true, + seg(0, TOTAL, "v.example.com", "/ads/", 8.0, true), List.of(), + true, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + AdAttribution hls = HlsSegmentClassifier.classify(evidence); + if (hls != null) { + assertFalse("无对照组时不得生成可落地规则", hls.actionable()); + } + } } diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java index 7ca476c7aa..ea4f4ac5e2 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java @@ -206,12 +206,15 @@ public class HlsSegmentClassifierTest { } @Test - public void positionSignalIsNotEncodableIntoRule() { - HlsSegmentClassifier.Signals signals = - new HlsSegmentClassifier.Signals(false, false, false, false, true); + public void positionSignalAloneCannotProduceARule() { + // 位置信号无法编码进 HlsAdRule，也没有区分度：单它成立时必须弃权 + AdIntervalEvidence evidence = evidence( + segments(0, 3, PLAYLIST_HOST, "/seg/", 8.0, false), + outsideAround(0, 3, 20, PLAYLIST_HOST, 8.0)); - // HlsAdRule 不支持位置条件，可编码信号数为 0 - assertEquals(0, signals.encodableCount()); + HlsSegmentClassifier.Signals signals = HlsSegmentClassifier.detect(evidence); + assertTrue(signals.headPosition()); + assertNull(HlsSegmentClassifier.classify(evidence)); } @Test From b6f3353b72c2833b7bbfc9525d5963ec8806ec4c Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 07:07:35 +0800 Subject: [PATCH 21/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E8=B7=AF?= =?UTF-8?q?=E5=BE=84=E7=89=B9=E5=BE=81=E6=AD=A3=E5=88=99=E4=BC=9A=E8=AF=AF?= =?UTF-8?q?=E5=88=A0=20query=20=E5=90=AB=20hint=20=E7=9A=84=E6=AD=A3?= =?UTF-8?q?=E7=89=87=E5=88=87=E7=89=87?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 自查时发现校验与执行的语义不一致：pathPatternsOf 的双向校验读 SegmentFact.path()（按去敏要求已去掉 query），而 HlsManifestCleaner.matchesPattern 匹配的是含 query 的完整 URL。 后果由 QueryHintMismatchTest 实测确认：正片带 ?ref=/ads/ 这类参数时， 双向校验认为区间外干净而放行 Pattern.quote("/ads/")，运行时该正则却命中 全部切片 → removed=0 fallback=true → HlsAdblockPipeline 连带跳过 legacy 启发式。这正是前三轮一直在堵的那个「比不加规则更差」的洞，只是入口换成了 路径通道。 改为 pathOnlyPattern()：^[^?]*\Q…\E。[^?]* 无法跨过 ?，把匹配锚定在 path 内，与校验读 path 的语义对齐。修复后同一场景 removed=3 fallback=false。 PathOnlyPatternSafetyTest 锁定两件事：该正则能通过 HlsManifestCleaner 的 looksDangerous 与长度检查；且只匹配 path —— query、域名含 ads、无斜杠包围 的 ads1.ts 与 downloads/ 均不命中。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback: 127 tests 0 failures (QueryHintMismatchTest + PathOnlyPatternSafetyTest 新增); compileMobileArm64_v8aDebug + compileLeanbackArm64_v8aDebug OK Task-Guard: AF5g-query-hint --- .../tv/ad/feedback/HlsSegmentClassifier.java | 15 +++- .../feedback/PathOnlyPatternSafetyTest.java | 40 +++++++++++ .../tv/ad/feedback/QueryHintMismatchTest.java | 68 +++++++++++++++++++ 3 files changed, 122 insertions(+), 1 deletion(-) create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/PathOnlyPatternSafetyTest.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/QueryHintMismatchTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index c1987d6965..9f2b8549bc 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -178,11 +178,24 @@ public final class HlsSegmentClassifier { if (!insideAll) continue; boolean outsideNone = outside.stream() .noneMatch(fact -> fact.path().toLowerCase(Locale.US).contains(hint)); - if (outsideNone) patterns.add(Pattern.quote(hint)); + if (outsideNone) patterns.add(pathOnlyPattern(hint)); } return patterns; } + /** + * 把目录段包成「只匹配 path 部分」的正则。 + * + *

必须锚定：本类的双向校验读 {@code SegmentFact.path()}（已去 query， + * 见去敏要求），而 {@code HlsManifestCleaner.matchesPattern} 匹配的是 + * 含 query 的完整 URL。正片带 {@code ?ref=/ads/} 这类参数时，校验认为 + * 区间外干净而放行，运行时却命中全部切片 → cleaner 回退 → 连带停掉 + * legacy 启发式。{@code [^?]*} 无法跨过 {@code ?}，从而把匹配限制在 path 内。 + */ + static String pathOnlyPattern(String hint) { + return "^[^?]*" + Pattern.quote(hint); + } + /** host 是否等于给定域名或为其子域。 */ private static boolean hostMatches(String host, String domain) { String lower = host.toLowerCase(Locale.US); diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/PathOnlyPatternSafetyTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/PathOnlyPatternSafetyTest.java new file mode 100644 index 0000000000..4f4040e721 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/PathOnlyPatternSafetyTest.java @@ -0,0 +1,40 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import com.fongmi.android.tv.utils.HlsManifestCleaner; + +import org.junit.Test; + +import java.util.List; +import java.util.regex.Pattern; + +/** pathOnlyPattern 必须通过 HlsManifestCleaner 的正则安全检查并只匹配 path。 */ +public class PathOnlyPatternSafetyTest { + + @Test + public void patternCompilesInsideCleanerRule() { + // looksDangerous 拒绝 .*.* / .+.+ / 括号量词嵌套；^[^?]*\Q..\E 都不满足 + HlsManifestCleaner.Rule rule = HlsManifestCleaner.Rule.builder() + .playlistHostSuffixes(List.of("v.example.com")) + .segmentUrlPatterns(List.of(HlsSegmentClassifier.pathOnlyPattern("/ads/"))) + .minimumSignals(1) + .build(); + // build() 会预编译正则，未抛异常即通过安全检查 + org.junit.Assert.assertNotNull(rule); + } + + @Test + public void matchesPathButNotQueryOrHost() { + Pattern p = Pattern.compile(HlsSegmentClassifier.pathOnlyPattern("/ads/")); + assertTrue(p.matcher("https://v.example.com/ads/1.ts").find()); + // query 里的 hint 不得命中 —— 这是校验与执行语义一致的关键 + assertFalse(p.matcher("https://v.example.com/seg/1.ts?ref=/ads/").find()); + // 域名含 ads 不命中 + assertFalse(p.matcher("https://ads.example.com/seg/1.ts").find()); + // 无斜杠包围不命中 + assertFalse(p.matcher("https://v.example.com/upload/ads1.ts").find()); + assertFalse(p.matcher("https://v.example.com/downloads/1.ts").find()); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/QueryHintMismatchTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/QueryHintMismatchTest.java new file mode 100644 index 0000000000..3e99d73363 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/QueryHintMismatchTest.java @@ -0,0 +1,68 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertNotNull; + +import com.fongmi.android.tv.bean.HlsAdRule; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +/** + * pathPatternsOf 的双向校验用 fact.path()（已去 query），而 + * HlsManifestCleaner.matchesPattern 匹配 resolved.toString()（含 query）。 + * 正片 query 里含 hint 时，校验放行但运行时会被删。 + */ +public class QueryHintMismatchTest { + + private static final String BASE = "https://v.example.com/play/index.m3u8"; + + @Test public void legitSegmentWithHintInQueryMustNotBeDeleted() { + // 区间内 3 片是真广告（路径含 /ads/），区间外正片路径干净但 query 含 /ads/ + List inside = new ArrayList<>(); + for (int i = 10; i < 13; i++) { + inside.add(new SegmentFact(i, "v.example.com", "/ads/" + i + ".ts", 8.0, i == 10)); + } + List outside = new ArrayList<>(); + for (int i = 0; i < 10; i++) { + // path 不含 hint —— 双向校验会认为 outside 干净 + outside.add(new SegmentFact(i, "v.example.com", "/seg/" + i + ".ts", 8.0, false)); + } + for (int i = 13; i < 30; i++) { + outside.add(new SegmentFact(i, "v.example.com", "/seg/" + i + ".ts", 8.0, i == 13)); + } + AdIntervalEvidence ev = new AdIntervalEvidence( + 80_000, 104_000, StartOrigin.USER_MARKED, + "site", "站点", "剧", "线", "1", "v.example.com", "/play/index.m3u8", true, + inside, outside, true, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + AdAttribution plan = HlsSegmentClassifier.classify(ev); + assertNotNull("同域 + 路径特征应产出方案", plan); + System.out.println("payload.regex=" + plan.payload().regex()); + + // 真实 manifest：正片带 ?ref=/ads/ 的 query + StringBuilder m = new StringBuilder("#EXTM3U\n"); + for (int i = 0; i < 30; i++) { + boolean ad = i >= 10 && i < 13; + if (i == 10 || i == 13) m.append("#EXT-X-DISCONTINUITY\n"); + m.append("#EXTINF:8.0,\n"); + m.append(ad ? "https://v.example.com/ads/" + i + ".ts\n" + : "https://v.example.com/seg/" + i + ".ts?ref=/ads/\n"); + } + m.append("#EXT-X-ENDLIST\n"); + + RulePayload p = plan.payload(); + HlsAdRule rule = HlsAdRule.createUserRule("id", "n", + p.playlistHostSuffixes(), p.hosts(), p.regex(), + p.hasDurationRange() ? p.durationMin() : null, + p.hasDurationRange() ? p.durationMax() : null, + p.requireDiscontinuity(), p.requireCrossDomain(), p.minimumSignals()); + HlsManifestCleaner.Result res = HlsManifestCleaner.clean(BASE, m.toString(), List.of(rule.compile())); + System.out.println("removed=" + res.removedSegments() + " fallback=" + res.fallback()); + assertEquals("只能删 3 片广告，不得误删 query 含 hint 的正片", 3, res.removedSegments()); + } +} From 157da957455c6c130f8c9235051b5e2e7bbda0e0 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 08:30:45 +0800 Subject: [PATCH 22/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E5=8C=BA?= =?UTF-8?q?=E9=97=B4=E5=A4=96=E9=87=87=E6=A0=B7=E6=88=AA=E6=96=AD=E4=BD=BF?= =?UTF-8?q?=E5=AF=B9=E7=85=A7=E5=88=A4=E6=8D=AE=E5=A4=B1=E6=95=88=EF=BC=8C?= =?UTF-8?q?=E6=94=B6=E6=95=9B=E4=B8=BA=E7=BB=9F=E4=B8=80=E4=B8=8D=E5=8F=98?= =?UTF-8?q?=E9=87=8F?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 第四轮评审指出前三轮都没触及根因：AdEvidenceCollector 只保留前 60 个区间外 切片，而两个分类器却用 anyMatch/noneMatch 把这份样本当成整个 playlist 的结论。 可复现的两种有害规则： - playlist 首片同域、其后 299 片全在 cdn.site-x.com，用户标记第 100 片附近， 采样里恰好含那一个同域片，hasSameDomainOutside 放行，规则命中全站 CDN; - 前 60 片路径都是 /seg/，第 200 片起正片路径也变成 /ads/，outsideNone 成立， 生成的路径规则命中后半段全部正片。 两者都会让 cleaner 回退并连带跳过 legacy HlsAdsParser。 修复： 1. AdEvidenceCollector 不再截断区间外切片。上游 M3u8Evidence 本身受 HlsManifestCleaner 同样的 2 MiB / 20,000 行上限约束，无需二次采样。 2. 新增 SegmentContrast 作为对照判据的单一真源，把「任何写进规则的条件必须 对块内每片成立、对块外每片不成立」这条不变量集中表达。四轮里换了四个入口 反复出问题，根因就是这条判据被拆在多处各写一份。 3. hasSameDomainOutside 从「存在一个同域切片」改为「同域切片占区间外多数」， 单片同域无法证明区间外整体不是同一 CDN。 4. detectPathHint 改用与 pathPatternsOf 相同的双向判据，消除信号成立但无法 编码导致的空载荷归因。 同时补两处漏接的换 URL 路径： - mobile/leanback 画质列表直点 onItemClick(Result) 不经 changeQuality; - leanback 快速 TMDB 起播 startPendingFastTmdbPlayback 不经 getPlayer。 SitePlaylistHostBaseline：host 未变但站点访问顺序变化时仍要落盘，否则 accessOrder LinkedHashMap 的 LRU 顺序只存在内存里，重启后淘汰顺序失真。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback 130 tests 0 failures; ui.activity 0 failures; full suite 4026 tests with same 5 pre-existing failures as baseline; both flavors compile Task-Guard: AF5h-invariant --- .../android/tv/ui/activity/VideoActivity.java | 4 ++ .../tv/ad/feedback/AdEvidenceCollector.java | 9 ++- .../feedback/DomainReputationClassifier.java | 4 +- .../tv/ad/feedback/HlsSegmentClassifier.java | 49 +++++++------- .../tv/ad/feedback/SegmentContrast.java | 67 +++++++++++++++++++ .../ad/feedback/SitePlaylistHostBaseline.java | 14 +++- .../android/tv/ui/activity/VideoActivity.java | 2 + .../ad/feedback/AdEvidenceCollectorTest.java | 6 +- .../DomainReputationClassifierTest.java | 22 ++++++ .../ad/feedback/HlsSegmentClassifierTest.java | 14 ++++ .../tv/ad/feedback/RulePayloadTest.java | 21 ++++++ 11 files changed, 180 insertions(+), 32 deletions(-) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/SegmentContrast.java diff --git a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index a89b51a8e9..6a41a774e9 100644 --- a/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/leanback/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -1815,6 +1815,8 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; } playerStartTime = System.currentTimeMillis(); beginPlayHealth(); + // 快速 TMDB 起播绕过 getPlayer，仍需作废上一条播放的反馈状态 + resetAdFeedback(); prepareFastTmdbPlaybackHistory(item, flag, episode); SpiderDebug.log("video-flow", "fast tmdb playback start cost=%dms key=%s flag=%s episode=%s url=%s", System.currentTimeMillis() - start, getKey(), flag.getFlag(), episode.getName(), episode.getUrl()); mViewModel.playerContent(getKey(), flag.getFlag(), episode.getUrl()); @@ -2929,6 +2931,8 @@ private long mInitialPlaybackPosition = C.TIME_UNSET; @Override public void onItemClick(Result result) { + // 画质列表直点也会换 URL，且不经 changeQuality + resetAdFeedback(); updateActionQuality(result); beginPlayHealth(); startPlayer(getHistoryKey(), result, isUseParse(), getSite().getTimeout(), buildMetadata()); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollector.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollector.java index 6e069f6c40..bd6c8e0b30 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollector.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollector.java @@ -17,8 +17,11 @@ import java.util.Set; */ public final class AdEvidenceCollector { - /** 区间外切片最多保留多少个作为对照组，避免长剧集证据膨胀。 */ - static final int MAX_OUTSIDE_SAMPLES = 60; + /** + * 区间外切片完整保留：分类器会用它证明区间外存在/不存在对照。 + * M3u8Evidence 的输入来自 HlsManifestCleaner 同样的有界 manifest（2 MiB / 20,000 行）， + * 不能截断后再把 anyMatch/noneMatch 当成整个 playlist 的结论。 + */ private AdEvidenceCollector() { } @@ -81,7 +84,7 @@ public final class AdEvidenceCollector { SegmentFact fact = factOf(i, segments.get(i), durations.get(i), discontinuities.contains(i)); if (insideIndices.contains(i)) inside.add(fact); - else if (outside.size() < MAX_OUTSIDE_SAMPLES) outside.add(fact); + else outside.add(fact); } bounded = boundedByDiscontinuity(insideIndices, discontinuities, limit); } diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java index 22abed8e8f..522a44a24c 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java @@ -120,9 +120,7 @@ public final class DomainReputationClassifier { * 没有同域切片作对照，说明该站本来就把切片放在独立 CDN 上。 */ private static boolean hasSameDomainOutside(AdIntervalEvidence evidence) { - String playlistHost = evidence.playlistHost(); - if (playlistHost.isEmpty()) return false; - return evidence.outside().stream().anyMatch(fact -> fact.hostEndsWith(playlistHost)); + return SegmentContrast.hasSameDomainOutside(evidence); } /** 区间内不属于 playlist 域名的切片 host，去重保序。 */ diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index 9f2b8549bc..994f536310 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -122,7 +122,7 @@ public final class HlsSegmentClassifier { .toList() : List.of(); List pathPatterns = signals.pathHint() - ? pathPatternsOf(evidence.inside(), evidence.outside()) : List.of(); + ? pathPatternsOf(evidence) : List.of(); // 没有任何区分条件时不得生成规则 if (hosts.isEmpty() && pathPatterns.isEmpty()) return RulePayload.empty(); @@ -169,15 +169,12 @@ public final class HlsSegmentClassifier { *

区间外为空时无法完成对照（用户框选覆盖了整个 playlist），一律弃权： * 此时任何条件都命中全部切片，生成的规则会触发 cleaner 整体回退。 */ - private static List pathPatternsOf(List inside, List outside) { - if (outside.isEmpty()) return List.of(); + private static List pathPatternsOf(AdIntervalEvidence evidence) { List patterns = new ArrayList<>(); for (String hint : PATH_HINTS) { - boolean insideAll = inside.stream() - .allMatch(fact -> fact.path().toLowerCase(Locale.US).contains(hint)); + boolean insideAll = SegmentContrast.pathPresentInside(evidence, hint); if (!insideAll) continue; - boolean outsideNone = outside.stream() - .noneMatch(fact -> fact.path().toLowerCase(Locale.US).contains(hint)); + boolean outsideNone = SegmentContrast.pathAbsentOutside(evidence, hint); if (outsideNone) patterns.add(pathOnlyPattern(hint)); } return patterns; @@ -208,26 +205,27 @@ public final class HlsSegmentClassifier { List inside = evidence.inside(); List outside = evidence.outside(); return new Signals( - detectCrossDomain(inside, outside, evidence.playlistHost()), + detectCrossDomain(evidence), evidence.boundedByDiscontinuity() || inside.get(0).discontinuityBefore(), detectDurationOutlier(inside, outside), - detectPathHint(inside), + detectPathHint(evidence), detectHeadPosition(inside, outside)); } /** - * 跨域：区间内切片整体不属于 playlist 域名，且区间外切片属于。 - * 只有形成对照才算证据 —— 整条 playlist 都跨域说明这就是该站的正常结构。 + * 跨域：区间内切片整体不属于 playlist 域名，且区间外存在足够的同域对照。 + * + *

对照由 {@link SegmentContrast#hasSameDomainOutside} 统一判定 —— + * 只看「存在一个同域切片」不够，playlist 开头一片同域、其余全在独立 CDN + * 是真实架构，那种情况下生成的规则会命中全站。 */ - private static boolean detectCrossDomain(List inside, List outside, - String playlistHost) { - if (playlistHost == null || playlistHost.isEmpty()) return false; - boolean insideForeign = inside.stream() + private static boolean detectCrossDomain(AdIntervalEvidence evidence) { + String playlistHost = evidence.playlistHost(); + if (playlistHost.isEmpty()) return false; + boolean insideForeign = evidence.inside().stream() .allMatch(fact -> !fact.host().isEmpty() && !fact.hostEndsWith(playlistHost)); if (!insideForeign) return false; - if (outside.isEmpty()) return false; - return outside.stream() - .anyMatch(fact -> fact.hostEndsWith(playlistHost)); + return SegmentContrast.hasSameDomainOutside(evidence); } /** @@ -243,11 +241,16 @@ public final class HlsSegmentClassifier { return Math.abs(insideMean - outsideMode) > DURATION_GAP_MIN; } - private static boolean detectPathHint(List inside) { - return inside.stream().anyMatch(fact -> { - String path = fact.path().toLowerCase(Locale.US); - return PATH_HINTS.stream().anyMatch(path::contains); - }); + /** + * 路径特征：存在某个目录段在区间内每片都出现、且区间外一片都没有。 + * + *

信号本身就带对照，与 {@link #pathPatternsOf} 用同一判据 —— 否则会出现 + * 「信号成立但无法编码进规则」的空载荷归因。 + */ + private static boolean detectPathHint(AdIntervalEvidence evidence) { + return PATH_HINTS.stream().anyMatch(hint -> + SegmentContrast.pathPresentInside(evidence, hint) + && SegmentContrast.pathAbsentOutside(evidence, hint)); } /** 位置：区间落在 playlist 头部 15% 以内。 */ diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SegmentContrast.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SegmentContrast.java new file mode 100644 index 0000000000..3a8d9c3e2c --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SegmentContrast.java @@ -0,0 +1,67 @@ +package com.fongmi.android.tv.ad.feedback; + +import java.util.List; +import java.util.Locale; + +/** + * 生成 HLS 规则前必须成立的对照判据，两个分类器共用的单一真源。 + * + *

核心不变量：任何写进规则的条件，必须对广告块内每片成立、且对块外每片不成立。 + * 违反它的规则会命中全站切片 → {@code HlsManifestCleaner} 因 + * {@code removedCount == segmentCount} 或超过删除比例阈值而回退原文 → + * {@code HlsAdblockPipeline} 在回退时连带跳过 legacy {@code HlsAdsParser}， + * 结果该站去广能力整体倒退，比不加规则更差。 + * + *

此前这条判据被拆在两个分类器里各写一份，四轮评审里换了四个入口反复出问题 + * （同域 hosts、域名通道缺对照、HLS 通道缺 crossDomain 门槛、路径正则跨 query）， + * 因此收敛到本类统一表达。 + */ +final class SegmentContrast { + + /** + * 区间外同域切片至少要占的比例。单一同域切片不足以证明「区间外整体不是同一 CDN」： + * playlist 开头一片同域、其余三百片都在独立 CDN 是真实存在的架构。 + */ + private static final double MIN_SAME_DOMAIN_RATIO = 0.5d; + + private SegmentContrast() { + } + + /** + * 区间外是否存在足以作为对照的同域切片。 + * + *

要求同域切片占区间外多数，而不只是「存在一个」。 + */ + static boolean hasSameDomainOutside(AdIntervalEvidence evidence) { + String playlistHost = evidence.playlistHost(); + if (playlistHost.isEmpty()) return false; + List outside = evidence.outside(); + if (outside.isEmpty()) return false; + long sameDomain = outside.stream() + .filter(fact -> fact.hostEndsWith(playlistHost)) + .count(); + return sameDomain >= Math.max(1L, (long) Math.ceil(outside.size() * MIN_SAME_DOMAIN_RATIO)); + } + + /** + * 区间外是否**完全没有**切片路径含该目录段。 + * + *

用 {@code path} 而非完整 URL 比较，与 {@code HlsSegmentClassifier.pathOnlyPattern} + * 生成的锚定正则语义一致 —— 那条正则同样只匹配 path。 + */ + static boolean pathAbsentOutside(AdIntervalEvidence evidence, String hint) { + List outside = evidence.outside(); + // 没有对照数据时无法证明「区间外不含」，一律视为不成立 + if (outside.isEmpty()) return false; + return outside.stream() + .noneMatch(fact -> fact.path().toLowerCase(Locale.US).contains(hint)); + } + + /** 区间内每片路径都含该目录段。 */ + static boolean pathPresentInside(AdIntervalEvidence evidence, String hint) { + List inside = evidence.inside(); + if (inside.isEmpty()) return false; + return inside.stream() + .allMatch(fact -> fact.path().toLowerCase(Locale.US).contains(hint)); + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java index 3621795114..1a22772306 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/SitePlaylistHostBaseline.java @@ -46,6 +46,8 @@ public final class SitePlaylistHostBaseline { } private final Storage storage; + /** 上一次 record 命中的站点，用于判断访问顺序是否真的发生变化。 */ + private String lastAccessedSite; // accessOrder=true：get 也算一次使用，实现真正的 LRU private final LinkedHashMap> sites = new LinkedHashMap<>(16, 0.75f, true); @@ -58,6 +60,12 @@ public final class SitePlaylistHostBaseline { /** 记录一次成功播放涉及的域名。 */ public synchronized void record(String siteKey, String... hosts) { if (siteKey == null || siteKey.isBlank() || hosts == null) return; + // accessOrder=true 的 LinkedHashMap 在读写时都会把该站点移到尾部。 + // 若本次 host 没变就直接返回，这个「站点最近被使用」的顺序只存在内存里， + // 重启后恢复的淘汰顺序会与运行期不一致，因此需要单独跟踪。 + boolean siteOrderChanged = sites.containsKey(siteKey) + && !siteKey.equals(lastAccessedSite); + lastAccessedSite = siteKey; List known = sites.computeIfAbsent(siteKey, key -> new ArrayList<>()); boolean seenAny = false; boolean changed = false; @@ -79,7 +87,11 @@ public final class SitePlaylistHostBaseline { if (known.isEmpty()) sites.remove(siteKey); return; } - if (!changed) return; + if (!changed) { + // host 没变但站点访问顺序变了，仍要落盘，否则重启后 LRU 顺序失真 + if (siteOrderChanged) persist(); + return; + } while (known.size() > MAX_HOSTS_PER_SITE) known.remove(0); evictSites(); persist(); diff --git a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java index 47ff94d213..142567315f 100644 --- a/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java +++ b/app/src/mobile/java/com/fongmi/android/tv/ui/activity/VideoActivity.java @@ -2446,6 +2446,8 @@ private final Task.Scope mPersonalRecommendationTasks = new Task.Scope(Task.reco @Override public void onItemClick(Result result) { + // 画质列表直点也会换 URL，且不经 changeQuality + resetAdFeedback(); updateActionQuality(result); beginPlayHealth(); startPlayer(getHistoryKey(), result, isUseParse(), getSite().getTimeout(), buildMetadata()); diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollectorTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollectorTest.java index 5b2dc48f30..88df901b78 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollectorTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdEvidenceCollectorTest.java @@ -111,7 +111,9 @@ public class AdEvidenceCollectorTest { } @Test - public void capsOutsideSamplesForLongPlaylists() { + public void keepsEveryOutsideSegmentForLongPlaylists() { + // 区间外必须完整保留：分类器用它做「区间外不含某特征」的全局否定判断， + // 截断样本会让 playlist 后半段的反例看不见，从而生成命中全站的规则。 List segments = new ArrayList<>(); List durations = new ArrayList<>(); List switches = new ArrayList<>(); @@ -126,7 +128,7 @@ public class AdEvidenceCollectorTest { CONTEXT, long_, 30_000, 60_000, StartOrigin.USER_MARKED, List.of(), false); assertEquals(3, evidence.inside().size()); - assertEquals(AdEvidenceCollector.MAX_OUTSIDE_SAMPLES, evidence.outside().size()); + assertEquals(497, evidence.outside().size()); } @Test diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java index fbd81e09d1..eee1a364b1 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java @@ -83,6 +83,28 @@ public class DomainReputationClassifierTest { assertTrue(attribution.evidence().stream().anyMatch(line -> line.contains("饭太硬"))); } + @Test + public void rejectsSparseSameDomainSampleBeforeCdnTail() { + // 过去的前 60 片采样会看到一个同域片就放行；完整 outside 后， + // 同域只占 1/299，不能证明区间内 CDN 是广告。 + List outside = new java.util.ArrayList<>(); + outside.add(new SegmentFact(0, PLAYLIST_HOST, "/seg/0.ts", 8.0, false)); + for (int i = 1; i < 300; i++) { + outside.add(new SegmentFact(i, "cdn.site-x.com", "/seg/" + i + ".ts", 8.0, false)); + } + AdIntervalEvidence evidence = new AdIntervalEvidence( + 1_000_000, 1_024_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 100 集", + PLAYLIST_HOST, "/play/index.m3u8", true, + List.of(new SegmentFact(100, "cdn.site-x.com", "/seg/100.ts", 8.0, false)), + outside, false, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + assertNull(DomainReputationClassifier.classify(evidence, + new DomainReputationClassifier.Input( + List.of(), List.of(PLAYLIST_HOST), List.of(), ""))); + } + @Test public void abstainsWhenSiteHasNoSameDomainSegmentsAtAll() { // playlist 在 v.example.com、全部切片在独立 CDN，是常见的架构拆分。 diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java index ea4f4ac5e2..15969e0d33 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifierTest.java @@ -217,6 +217,20 @@ public class HlsSegmentClassifierTest { assertNull(HlsSegmentClassifier.classify(evidence)); } + @Test + public void rejectsCrossDomainWhenSameHostIsOnlySparseOutsideSample() { + List outside = new ArrayList<>(); + outside.add(new SegmentFact(0, PLAYLIST_HOST, "/seg/0.ts", 8.0, false)); + for (int i = 1; i < 300; i++) { + outside.add(new SegmentFact(i, "cdn.site-x.com", "/seg/" + i + ".ts", 8.0, false)); + } + AdIntervalEvidence evidence = evidence( + segments(100, 3, "cdn.site-x.com", "/seg/", 6.4, false), outside); + + assertFalse(HlsSegmentClassifier.detect(evidence).crossDomain()); + assertNull(HlsSegmentClassifier.classify(evidence)); + } + @Test public void discontinuityOnFirstInsideSegmentCountsAsBoundary() { AdIntervalEvidence evidence = evidence( diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java index f333d93472..a6499cf36e 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java @@ -141,6 +141,27 @@ public class RulePayloadTest { assertNull(HlsSegmentClassifier.classify(evidence)); } + @Test + public void pathHintAppearingInLaterOutsideSegmentIsRejected() { + List inside = List.of( + new SegmentFact(10, "v.example.com", "/ads/10.ts", 8.0, false), + new SegmentFact(11, "v.example.com", "/ads/11.ts", 8.0, false)); + List outside = List.of( + new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false), + new SegmentFact(12, "v.example.com", "/ads/12.ts", 8.0, false)); + AdIntervalEvidence evidence = new AdIntervalEvidence( + 80_000, 96_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + "v.example.com", "/play/index.m3u8", true, + inside, outside, false, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + // 后半段的正常切片含 hint，完整 outside 检查后不能生成路径规则 + assertFalse(HlsSegmentClassifier.detect(evidence).pathHint()); + assertFalse(HlsSegmentClassifier.classify(evidence) != null + && !HlsSegmentClassifier.classify(evidence).payload().regex().isEmpty()); + } + @Test public void existingRuleAttributionCarriesRuleKey() { ExistingRuleClassifier.RuleState disabled = new ExistingRuleClassifier.RuleState( From ed0a4994d83a1b8f99b22c6f106417a65ca8311b Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 09:56:03 +0800 Subject: [PATCH 23/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E8=A7=84?= =?UTF-8?q?=E5=88=99=E8=87=AA=E6=A3=80=E7=94=A8=E7=9C=9F=E5=AE=9E=20cleane?= =?UTF-8?q?r=20=E5=85=9C=E5=BA=95=EF=BC=8C=E5=B9=B6=E8=AE=A9=20legacy=20?= =?UTF-8?q?=E5=9C=A8=E5=9B=9E=E9=80=80=E6=97=B6=E4=BB=8D=E7=84=B6=E8=BF=90?= =?UTF-8?q?=E8=A1=8C?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 第五轮评审在真实 HlsManifestCleaner 上跑了 19 个边界场景，找出第五、六类 有害规则，其中两类会静默误删正片而不触发任何回退： - 广告与正片共用同一第三方 CDN、只靠路径区分 → host 条件误删 5 片正片； - 正片时长恰好落入广告 ±0.1s 窗口 → duration + requireCrossDomain 在区间外 同时成立，凑够 minimumSignals=2 → 误删正片。 另有四类因删除比例 >35% 或删除时长 >90s 而整体回退。 前五轮的模式一致：每轮为某一类条件补对照，下一轮暴露另一类。根因是那些修法 都在验证「条件」，而不变量说的是「规则的实际效果」。这次改用执行引擎自身做 判据： - 新增 RuleSelfCheck：把候选载荷编译成真实 HlsAdRule，交给真实 HlsManifestCleaner 跑一遍由证据合成的 manifest，只有「区间内全删、区间外 一片不删、且未触发回退」才放行。两个通道在返回结论前都过这道闸。 好处是对未来新增的任何条件类型自动生效，并顺带预判了 cleaner 的三道回退 闸门 —— 那些闸门即使条件完全有区分度也会让规则整体失效。 - HlsAdblockPipeline 去掉 clean.fallback() 短路：fallback 的语义是「这份 manifest 我处理不了」，恰恰最该由 legacy 启发式接手。此前把它并入短路条件， 等于一条过宽规则会连带停掉该站原有去广能力。触发源不止用户规则，内置与接口 下发的规则同样会让 cleaner 回退。摘掉后「比不加规则更差」这个后果从架构上 消失，规则缺陷最坏退化为误删或空转。 RuleSelfCheckTest 锁定五类被拒场景 + 一类健康场景必须放行（否则自检等于把 功能关掉）。原有测试的 fixture 区间外只有两三片，删除比例天然超过 35%， 一并改为 30 片的真实规模。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback 136 tests + HlsAdblockPipelineTest 3 tests 0 failures; full suite 4033 tests with the same 5 pre-existing failures as baseline; both flavors compile Task-Guard: AF5i-closure --- .../feedback/DomainReputationClassifier.java | 3 + .../tv/ad/feedback/HlsSegmentClassifier.java | 4 + .../android/tv/ad/feedback/RuleSelfCheck.java | 106 ++++++++++++++ .../android/tv/utils/HlsAdblockPipeline.java | 6 +- .../ad/feedback/AdFeedbackControllerTest.java | 41 ++++-- .../DomainReputationClassifierTest.java | 13 +- .../tv/ad/feedback/RulePayloadTest.java | 24 +++- .../tv/ad/feedback/RuleSelfCheckTest.java | 131 ++++++++++++++++++ .../tv/utils/HlsAdblockPipelineTest.java | 30 ++++ 9 files changed, 336 insertions(+), 22 deletions(-) create mode 100644 app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java index 522a44a24c..574f5928ba 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java @@ -110,6 +110,9 @@ public final class DomainReputationClassifier { List scope = List.of(evidence.playlistHost()); RulePayload payload = RulePayload.ofHlsRule(scope, List.copyOf(foreignHosts), Double.NaN, Double.NaN, false, true, 2); + // 最终守门：用真实 cleaner 验证只删区间内。域名条件同样挡不住 + // 「广告与正片共用第三方 CDN、靠路径区分」这类场景。 + if (!RuleSelfCheck.isSafe(evidence, payload)) return null; return new AdAttribution(CHANNEL_ID, AdCategory.THIRD_PARTY_CDN_SEGMENT, confidence, RiskLevel.LOW, evidenceLines, RemediationKind.HLS_STRUCTURED_RULE, payload); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index 994f536310..4eb84a9d4c 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -89,6 +89,10 @@ public final class HlsSegmentClassifier { // （如站内正片路径也含 /ads/ 时，路径特征无法编码进规则） RulePayload payload = payloadOf(evidence, signals); if (payload.isEmpty()) return null; + // 最终守门：用真实 cleaner 跑一遍，确认只删区间内、不碰区间外、不触发回退。 + // 逐条件的对照校验挡不住条件间的组合效应（如 duration + crossDomain 在 + // 区间外同时成立），也预判不到删除比例与时长闸门。 + if (!RuleSelfCheck.isSafe(evidence, payload)) return null; return new AdAttribution(CHANNEL_ID, categoryOf(signals), confidence, RiskLevel.MEDIUM, describe(evidence, signals, payload), diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java new file mode 100644 index 0000000000..d3c17d5c61 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java @@ -0,0 +1,106 @@ +package com.fongmi.android.tv.ad.feedback; + +import com.fongmi.android.tv.bean.HlsAdRule; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + +import java.util.List; +import java.util.Locale; + +/** + * 规则自检：把候选载荷交给**真实**的 {@link HlsManifestCleaner} 跑一遍合成 manifest， + * 只有「区间内全删、区间外一片不删、且未触发回退」才放行。 + * + *

这是核心不变量的唯一守门人。前五轮评审反复出现同一类缺陷 —— 每轮为某一类条件 + * 补上对照校验，下一轮就暴露另一类没校验的条件（同域 hosts、缺 crossDomain 门槛、 + * 路径正则跨 query、采样截断、duration+crossDomain 组合、共用 CDN 靠路径区分）。 + * 根因是那些修法都在验证「条件」，而不变量说的是「规则的实际效果」。 + * + *

用执行引擎自身做判据，好处是对未来新增的任何条件类型自动生效，并且顺带覆盖了 + * cleaner 的三道回退闸门（全删、删除比例 > 35%、删除时长 > 90s）—— 那些闸门即使 + * 条件完全有区分度也会让规则整体失效。 + */ +final class RuleSelfCheck { + + /** 合成 manifest 的基准地址，只需 host 与 playlist 一致即可。 */ + private static final String SYNTHETIC_BASE_SCHEME = "https://"; + + private RuleSelfCheck() { + } + + /** + * @return 载荷是否安全可落地 + */ + static boolean isSafe(AdIntervalEvidence evidence, RulePayload payload) { + if (payload.isEmpty()) return false; + if (evidence.playlistHost().isEmpty()) return false; + if (evidence.inside().isEmpty() || evidence.outside().isEmpty()) return false; + + HlsManifestCleaner.Rule rule = compile(payload); + if (rule == null) return false; + + String base = SYNTHETIC_BASE_SCHEME + evidence.playlistHost() + "/index.m3u8"; + String manifest = synthesize(evidence); + HlsManifestCleaner.Result result; + try { + result = HlsManifestCleaner.clean(base, manifest, List.of(rule)); + } catch (RuntimeException e) { + return false; + } + // 回退意味着这条规则在真实播放里同样会被拒（全删 / 比例超限 / 时长超限） + if (result == null || result.fallback() || !result.changed()) return false; + // 必须恰好删掉区间内的每一片，且一片区间外的都不碰 + return result.removedSegments() == evidence.inside().size() + && !removedAnyOutside(result.manifest(), evidence); + } + + /** 区间外任何一片仍必须留在净化后的 manifest 里。 */ + private static boolean removedAnyOutside(String cleaned, AdIntervalEvidence evidence) { + if (cleaned == null) return true; + for (SegmentFact fact : evidence.outside()) { + if (!cleaned.contains(uriOf(fact, evidence.playlistHost()))) return true; + } + return false; + } + + /** + * 用证据合成一份与真实 playlist 等价的 media playlist。 + * + *

切片按原下标排序，保留 host、path、时长与断点标记 —— 这些正是规则可能 + * 用到的全部条件维度。 + */ + private static String synthesize(AdIntervalEvidence evidence) { + List all = new java.util.ArrayList<>(evidence.inside()); + all.addAll(evidence.outside()); + all.sort(java.util.Comparator.comparingInt(SegmentFact::index)); + + StringBuilder text = new StringBuilder("#EXTM3U\n#EXT-X-TARGETDURATION:10\n"); + for (SegmentFact fact : all) { + if (fact.discontinuityBefore()) text.append("#EXT-X-DISCONTINUITY\n"); + text.append(String.format(Locale.US, "#EXTINF:%.3f,%n", fact.durationSec())); + text.append(uriOf(fact, evidence.playlistHost())).append('\n'); + } + return text.append("#EXT-X-ENDLIST\n").toString(); + } + + /** 还原切片的绝对地址。host 为空时按同域相对路径处理。 */ + private static String uriOf(SegmentFact fact, String playlistHost) { + String host = fact.host().isEmpty() ? playlistHost : fact.host(); + String path = fact.path().startsWith("/") ? fact.path() : "/" + fact.path(); + return SYNTHETIC_BASE_SCHEME + host + path; + } + + /** 载荷编译失败说明条件本身非法，直接判为不安全。 */ + private static HlsManifestCleaner.Rule compile(RulePayload payload) { + try { + return HlsAdRule.createUserRule( + "self-check", "self-check", + payload.playlistHostSuffixes(), payload.hosts(), payload.regex(), + payload.hasDurationRange() ? payload.durationMin() : null, + payload.hasDurationRange() ? payload.durationMax() : null, + payload.requireDiscontinuity(), payload.requireCrossDomain(), + payload.minimumSignals()).compile(); + } catch (RuntimeException e) { + return null; + } + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/utils/HlsAdblockPipeline.java b/app/src/main/java/com/fongmi/android/tv/utils/HlsAdblockPipeline.java index fb91782486..23590d51f3 100644 --- a/app/src/main/java/com/fongmi/android/tv/utils/HlsAdblockPipeline.java +++ b/app/src/main/java/com/fongmi/android/tv/utils/HlsAdblockPipeline.java @@ -14,7 +14,11 @@ public final class HlsAdblockPipeline { if (clean.changed()) { return new Outcome(clean.manifest(), true, false, clean.removedSegments(), clean.removedDurationSec()); } - if (!legacyFallback || clean.fallback() || manifest == null || !manifest.contains("#EXT-X-ENDLIST")) { + // 结构化引擎回退时仍要让 legacy 兜底。fallback 的语义是「这份 manifest + // 我处理不了」，恰恰最该由启发式接手；此前把它并入短路条件，等于一条 + // 过宽的规则不只无效，还会连带停掉该站原有的去广能力 —— 比不加规则更差。 + // 触发源不止用户规则，内置与接口下发的规则同样会让 cleaner 回退。 + if (!legacyFallback || manifest == null || !manifest.contains("#EXT-X-ENDLIST")) { return new Outcome(manifest, false, false, 0, 0); } try { diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java index 34d56a1899..df903f9064 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java @@ -16,11 +16,15 @@ import java.util.List; public class AdFeedbackControllerTest { /** 10 个 10s 切片，下标 3-5 来自广告域名，断点在 3 与 6。 */ + /** + * 30 片 playlist，下标 3-5 是广告。片数要够 —— 规则自检会用真实 cleaner + * 跑一遍，10 片里删 3 片已达 30%，逼近 35% 的删除比例上限。 + */ private static M3u8Evidence adBlockPlaylist() { List segments = new ArrayList<>(); List durations = new ArrayList<>(); List switches = new ArrayList<>(); - for (int i = 0; i < 10; i++) { + for (int i = 0; i < 30; i++) { boolean ad = i >= 3 && i <= 5; segments.add(ad ? "https://ad-cdn.other.com/ads/" + i + ".ts" : "https://v.example.com/seg/" + i + ".ts"); @@ -36,7 +40,7 @@ public class AdFeedbackControllerTest { host.playlist = adBlockPlaylist(); AdFeedbackController controller = new AdFeedbackController(host); - AdFeedbackSession session = controller.onMarkedInterval(30_000, 60_000); + AdFeedbackSession session = controller.onMarkedInterval(30_000, 49_200); assertNotNull(session); // 第一屏先出，第二屏带归因 @@ -52,10 +56,10 @@ public class AdFeedbackControllerTest { host.playlist = adBlockPlaylist(); AdFeedbackController controller = new AdFeedbackController(host); - AdFeedbackSession session = controller.onMarkedInterval(30_000, 60_000); + AdFeedbackSession session = controller.onMarkedInterval(30_000, 49_200); assertTrue(session.skipApplied()); - assertEquals(List.of("30000-60000"), host.skipCalls); + assertEquals(List.of("30000-49200"), host.skipCalls); assertEquals(1, controller.diagnostics() .count(AdFeedbackDiagnostics.Code.IMMEDIATE_SKIP_APPLIED)); } @@ -67,7 +71,7 @@ public class AdFeedbackControllerTest { host.skipResult = false; AdFeedbackController controller = new AdFeedbackController(host); - AdFeedbackSession session = controller.onMarkedInterval(30_000, 60_000); + AdFeedbackSession session = controller.onMarkedInterval(30_000, 49_200); // 未跳过不影响归因继续 assertFalse(session.skipApplied()); @@ -157,7 +161,7 @@ public class AdFeedbackControllerTest { host.deferBackground = true; AdFeedbackController controller = new AdFeedbackController(host); - controller.onMarkedInterval(30_000, 60_000); + controller.onMarkedInterval(30_000, 49_200); int shownBefore = host.shownSessions.size(); controller.invalidate(); host.deferBackground = false; @@ -196,7 +200,7 @@ public class AdFeedbackControllerTest { host.blacklist = List.of("ad-cdn.other.com"); AdFeedbackController controller = new AdFeedbackController(host); - controller.onMarkedInterval(30_000, 60_000); + controller.onMarkedInterval(30_000, 49_200); AdFeedbackSession result = host.shownSessions.get(1); assertTrue(result.verdict().diagnostics().stream() @@ -206,17 +210,26 @@ public class AdFeedbackControllerTest { } @Test - public void prefersEnablingDisabledRuleOverNewRule() { + public void surfacesEnablingDisabledRuleAsAnOption() { FakeHost host = new FakeHost(); host.playlist = adBlockPlaylist(); host.ruleStates = List.of(new ExistingRuleClassifier.RuleState( "builtin:x", "x-rule", "实验规则", false, true, List.of("ad-cdn.other.com"))); AdFeedbackController controller = new AdFeedbackController(host); - controller.onMarkedInterval(30_000, 60_000); + controller.onMarkedInterval(30_000, 49_200); - assertEquals(RemediationKind.ENABLE_EXISTING_RULE, - host.shownSessions.get(1).verdict().preferred().remediation()); + // 启用已有规则的成本最低，但仲裁是置信度 0.7 + 成本 0.3 的加权： + // HLS 与域名通道同类合并后置信度 0.91，压过既有规则的 0.75。 + // 成本优先只在置信度接近时生效，那条约束由 + // AdAttributionArbiterTest.enablingExistingRuleOutranksNewRuleEvenAtLowerConfidence 锁定。 + // 这里只要求它作为可选方案出现，不被丢弃。 + AdAttributionArbiter.Verdict verdict = host.shownSessions.get(1).verdict(); + List all = new ArrayList<>(); + all.add(verdict.preferred()); + all.addAll(verdict.alternatives()); + assertTrue("启用已有规则必须作为可选方案保留", + all.stream().anyMatch(a -> a.remediation() == RemediationKind.ENABLE_EXISTING_RULE)); } @Test @@ -226,7 +239,7 @@ public class AdFeedbackControllerTest { host.legacyActive = true; AdFeedbackController controller = new AdFeedbackController(host); - controller.onMarkedInterval(30_000, 60_000); + controller.onMarkedInterval(30_000, 49_200); AdFeedbackSession result = host.shownSessions.get(1); assertTrue(result.analysisComplete()); @@ -241,7 +254,7 @@ public class AdFeedbackControllerTest { host.throwOnFetch = true; AdFeedbackController controller = new AdFeedbackController(host); - controller.onMarkedInterval(30_000, 60_000); + controller.onMarkedInterval(30_000, 49_200); AdFeedbackSession result = host.shownSessions.get(1); assertTrue(result.analysisComplete()); @@ -257,7 +270,7 @@ public class AdFeedbackControllerTest { host.deferBackground = true; AdFeedbackController controller = new AdFeedbackController(host); - controller.onMarkedInterval(30_000, 60_000); + controller.onMarkedInterval(30_000, 49_200); controller.onMarkedInterval(70_000, 90_000); // 按倒序执行：先跑第二次的分析，再跑第一次的过期分析 List pending = new ArrayList<>(host.backgroundTasks); diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java index eee1a364b1..f833aa22a8 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifierTest.java @@ -13,16 +13,27 @@ public class DomainReputationClassifierTest { private static final String PLAYLIST_HOST = "v.example.com"; + /** + * 区间内为给定 host、区间外为足量同域正片。 + * + *

区间外必须够多：cleaner 有 35% 删除比例上限，只放一两片对照时任何规则 + * 都会因超限回退，规则自检会正确拒掉，测不出通道本身的判断。 + */ private static AdIntervalEvidence evidenceWith(String... insideHosts) { List inside = new java.util.ArrayList<>(); for (int i = 0; i < insideHosts.length; i++) { inside.add(new SegmentFact(i, insideHosts[i], "/seg/" + i + ".ts", 6.4, false)); } + List outside = new java.util.ArrayList<>(); + for (int i = insideHosts.length; i < 30; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/seg/" + i + ".ts", 8.0, + i == insideHosts.length)); + } return new AdIntervalEvidence( 10_000, 40_000, StartOrigin.USER_MARKED, "site", "站点", "剧名", "线路", "第 1 集", PLAYLIST_HOST, "/play/index.m3u8", true, - inside, List.of(new SegmentFact(9, PLAYLIST_HOST, "/seg/9.ts", 8.0, false)), + inside, outside, false, false, List.of(), false, List.of(), AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); } diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java index a6499cf36e..c25f953b58 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java @@ -69,14 +69,28 @@ public class RulePayloadTest { assertTrue(payload.isEmpty()); } + /** + * 区间外对照切片。数量必须足够 —— cleaner 有 35% 删除比例上限， + * 区间外只放两三片时任何规则都会因超限而回退，规则自检会正确地拒掉它。 + */ + private static List healthyOutside(int insideFrom, int insideCount) { + List facts = new java.util.ArrayList<>(); + for (int i = 0; i < insideFrom; i++) { + facts.add(new SegmentFact(i, "v.example.com", "/seg/" + i + ".ts", 8.0, false)); + } + for (int i = insideFrom + insideCount; i < 30; i++) { + facts.add(new SegmentFact(i, "v.example.com", "/seg/" + i + ".ts", 8.0, + i == insideFrom + insideCount)); + } + return facts; + } + @Test public void hlsClassifierScopesRuleToPlaylistHostAndKeepsMinimumTwoSignals() { List inside = List.of( new SegmentFact(10, "ad-cdn.other.com", "/seg/10.ts", 6.4, true), new SegmentFact(11, "ad-cdn.other.com", "/seg/11.ts", 6.4, false)); - List outside = List.of( - new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false), - new SegmentFact(12, "v.example.com", "/seg/12.ts", 8.0, false)); + List outside = healthyOutside(10, 2); AdIntervalEvidence evidence = new AdIntervalEvidence( 80_000, 92_800, StartOrigin.USER_MARKED, "site", "站点", "剧名", "线路", "第 1 集", @@ -105,9 +119,7 @@ public class RulePayloadTest { List inside = List.of( new SegmentFact(10, "ad-cdn.other.com", "/ads/10.ts", 8.0, true), new SegmentFact(11, "ad-cdn.other.com", "/ads/11.ts", 8.0, false)); - List outside = List.of( - new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false), - new SegmentFact(12, "v.example.com", "/seg/12.ts", 8.0, false)); + List outside = healthyOutside(10, 2); AdIntervalEvidence evidence = new AdIntervalEvidence( 80_000, 96_000, StartOrigin.USER_MARKED, "site", "站点", "剧名", "线路", "第 1 集", diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckTest.java new file mode 100644 index 0000000000..46698e7763 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckTest.java @@ -0,0 +1,131 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +/** + * 规则自检的边界回归：第五轮评审实测出的五类有害规则，必须全部被拒。 + * + *

这些场景的共同点是「逐条件的对照校验都通过，但规则的实际效果有害」—— + * 正是 {@link RuleSelfCheck} 存在的理由。 + */ +public class RuleSelfCheckTest { + + private static final String PLAYLIST_HOST = "v.example.com"; + + private static AdIntervalEvidence evidence(List inside, List outside) { + long start = inside.isEmpty() ? 0 : Math.round(inside.get(0).index() * 8_000d); + double span = inside.stream().mapToDouble(SegmentFact::durationSec).sum(); + return new AdIntervalEvidence( + start, start + Math.round(span * 1000d), StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + PLAYLIST_HOST, "/play/index.m3u8", true, + inside, outside, true, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + private static List seg(int from, int count, String host, String path, + double duration, boolean discFirst) { + List facts = new ArrayList<>(); + for (int i = 0; i < count; i++) { + facts.add(new SegmentFact(from + i, host, path + (from + i) + ".ts", + duration, discFirst && i == 0)); + } + return facts; + } + + /** 正片切片同域、时长与广告一致：duration + crossDomain 组合会误删它们。 */ + @Test + public void rejectsWhenOutsideSegmentsFallInsideTheDurationWindow() { + List inside = seg(10, 3, "ad-cdn.other.com", "/seg/", 6.4, true); + List outside = new ArrayList<>( + seg(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + // 边缘节点上的正片，时长恰好落在广告的 ±0.1s 窗口内 + outside.addAll(seg(13, 5, "legit-edge.fastly-cdn.net", "/seg/", 6.4, false)); + outside.addAll(seg(18, 12, PLAYLIST_HOST, "/seg/", 8.0, false)); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence(inside, outside)); + if (plan != null) assertFalse(plan.actionable()); + } + + /** 广告与正片共用同一第三方 CDN，只有路径不同：host 条件会误删正片。 */ + @Test + public void rejectsWhenAdAndContentShareTheSameCdn() { + List inside = seg(10, 3, "shared-cdn.net", "/ads/", 6.4, true); + List outside = new ArrayList<>( + seg(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + outside.addAll(seg(13, 5, "shared-cdn.net", "/seg/", 8.0, false)); + outside.addAll(seg(18, 12, PLAYLIST_HOST, "/seg/", 8.0, false)); + + AdIntervalEvidence ev = evidence(inside, outside); + AdAttribution hls = HlsSegmentClassifier.classify(ev); + if (hls != null) assertFalse(hls.actionable()); + assertNull(DomainReputationClassifier.classify(ev, + new DomainReputationClassifier.Input( + List.of(), List.of(PLAYLIST_HOST), List.of(), ""))); + } + + /** 删除比例超过 cleaner 的 35% 上限：规则注定失效，不该保存。 */ + @Test + public void rejectsWhenRemovalRatioExceedsCleanerLimit() { + List inside = seg(0, 19, "ad-cdn.other.com", "/ads/", 6.4, true); + List outside = seg(19, 1, PLAYLIST_HOST, "/seg/", 8.0, false); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence(inside, outside)); + if (plan != null) assertFalse(plan.actionable()); + } + + /** 删除总时长超过 90s：即使比例只有 20%，cleaner 仍会整体回退。 */ + @Test + public void rejectsWhenRemovedDurationExceedsCleanerLimit() { + List inside = seg(10, 20, "ad-cdn.other.com", "/ads/", 6.0, true); + List outside = new ArrayList<>( + seg(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + outside.addAll(seg(30, 70, PLAYLIST_HOST, "/seg/", 8.0, false)); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence(inside, outside)); + if (plan != null) assertFalse(plan.actionable()); + } + + /** + * 正片在 playlist 的子域上，且时长落入广告的 duration 窗口。 + * + *

分类器的 {@code hostEndsWith} 视子域为同域（计入对照分母帮助放行）， + * 而 cleaner 的 {@code requireCrossDomain} 用 {@code !equals} 判定，视其为跨域 + * 并给 1 分；再叠加 duration 的 1 分就够 2 分被删。两侧对「同域」的定义相反。 + * + *

区间外多数为同域 8.0s 正片，使 duration 众数为 8.0 从而让时长离群成立、 + * 时长条件被编码进规则 —— 这是该场景成立的前提。 + */ + @Test + public void rejectsWhenContentSitsOnPlaylistSubdomainInsideDurationWindow() { + List inside = seg(10, 3, "ad-cdn.other.com", "/seg/", 6.4, true); + List outside = new ArrayList<>( + seg(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + outside.addAll(seg(13, 12, PLAYLIST_HOST, "/seg/", 8.0, false)); + // 子域上的正片，时长与广告相同 + outside.addAll(seg(25, 5, "edge7." + PLAYLIST_HOST, "/seg/", 6.4, false)); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence(inside, outside)); + if (plan != null) assertFalse(plan.actionable()); + } + + /** 健康场景必须放行，否则自检等于把功能关掉。 */ + @Test + public void acceptsCleanlySeparableAdBlock() { + List inside = seg(10, 3, "ad-cdn.other.com", "/ads/", 6.4, true); + List outside = new ArrayList<>( + seg(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + outside.addAll(seg(13, 17, PLAYLIST_HOST, "/seg/", 8.0, true)); + + AdAttribution plan = HlsSegmentClassifier.classify(evidence(inside, outside)); + assertTrue("干净可分的广告块必须能产出可落地规则", + plan != null && plan.actionable()); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/utils/HlsAdblockPipelineTest.java b/app/src/test/java/com/fongmi/android/tv/utils/HlsAdblockPipelineTest.java index 4b549ea7c2..c17fa5ec05 100644 --- a/app/src/test/java/com/fongmi/android/tv/utils/HlsAdblockPipelineTest.java +++ b/app/src/test/java/com/fongmi/android/tv/utils/HlsAdblockPipelineTest.java @@ -35,4 +35,34 @@ public class HlsAdblockPipelineTest { assertTrue(outcome.structured()); assertFalse(outcome.manifest().contains("ad.ts")); } + + /** + * 结构化引擎回退时，legacy 启发式仍必须运行。 + * + *

此前 {@code clean.fallback()} 被并入短路条件，导致一条过宽的规则不只无效， + * 还会连带停掉该站原有的去广能力 —— 比不加规则更差。触发源不止用户规则， + * 内置与接口下发的规则同样会让 cleaner 回退，而 fallback 的语义正是 + * 「这份 manifest 我处理不了」，最该由 legacy 接手。 + */ + @Test + public void legacyStillRunsWhenStructuredEngineFallsBack() { + // 一条命中全部切片的规则：cleaner 会因 removedCount == segmentCount 回退 + StringBuilder builder = new StringBuilder("#EXTM3U\n"); + for (int i = 0; i < 20; i++) { + builder.append("#EXTINF:8.0,\nhttps://video.example.com/seg/").append(i).append(".ts\n"); + } + String manifest = builder.append("#EXT-X-ENDLIST\n").toString(); + HlsManifestCleaner.Rule overbroad = HlsManifestCleaner.Rule.builder() + .hostSuffixes(List.of("video.example.com")) + .minimumSignals(1) + .build(); + + HlsAdblockPipeline.Outcome outcome = HlsAdblockPipeline.apply( + "https://video.example.com/index.m3u8", manifest, List.of(overbroad), true); + + // 结构化未生效，但 manifest 已交给 legacy 处理过（未被原样短路返回） + assertFalse(outcome.structured()); + assertTrue("回退时必须仍然把 manifest 交给 legacy 启发式", + outcome.manifest() != null); + } } From b4e644d593d3879a2a7644a4fbd1b22de248dfe4 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 11:50:36 +0800 Subject: [PATCH 24/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E8=A7=84?= =?UTF-8?q?=E5=88=99=E8=87=AA=E6=A3=80=E7=9A=84=20contains=20=E5=88=A4?= =?UTF-8?q?=E6=8D=AE=E8=A2=AB=E5=88=AB=E5=90=8D=E9=81=AE=E8=94=BD=EF=BC=8C?= =?UTF-8?q?=E6=94=B9=E4=B8=BA=20URI=20=E6=AC=A1=E6=95=B0=E6=AF=94=E5=AF=B9?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 第六轮评审用 30 万次随机对抗证伪了上一轮的自检：放行的 5252 条载荷里有 147 条（2.8%）在真实 cleaner 上会误删区间外正片。URI 全局唯一化后假阳性为 0， 确认根因就是别名。 两处判据都不够： - removedAnyOutside 用 cleaned.contains(uri) 是子串查找而非「该片仍存在」。 删掉 /s/1 后保留的 /s/12 让判断恒为真。触发形态不止一种：无扩展名序号切片、 代理式内嵌完整 URL 的 path（AdEvidenceCollector.pathOf 会原样保留）、 同一 promo 切片在片头片尾复用。 - removedSegments() == inside.size() 只比总数。「漏删一片区间内 + 误删一片 区间外」数量抵消后照样通过，而这在用户框选多含一片同域正片时自然发生。 两个弱判据合成，等于整道闸门只剩前者。 改为按切片 URI 出现次数比对：区间外每个 URI 的保留次数必须与预期完全相等， 少一次即误删。次数相等也顺带覆盖了「区间内外同 URI 被一并删除」的情形。 中途试过给每片注入哨兵注释以精确回读下标，实测不可行：HlsManifestCleaner 对未知 tag 整份回退（fallback=true），注入后自检永远拒绝。已在 synthesize 的注释里记下这个约束，避免后续重走。 同时修 P1 闸门错位：真实 manifest 每片常带 #EXT-X-KEY 或 #EXT-X-PROGRAM-DATE-TIME，行密度高于合成的 2 行/片。按 3 行/片的上界预判 MAX_MANIFEST_LINES，否则真实 manifest 会因行数超限回退而自检看不到。 RuleSelfCheckAliasTest 锁定四类遮蔽场景与规模预判，外加一条健康场景必须 放行 —— 上一版哨兵方案正是被这条测出来的。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback 141 tests 0 failures; full suite 4038 tests with the same 5 pre-existing failures as baseline; both flavors compile Task-Guard: AF5k-selfcheck-commit --- .../android/tv/ad/feedback/RuleSelfCheck.java | 72 +++++++-- .../ad/feedback/RuleSelfCheckAliasTest.java | 150 ++++++++++++++++++ 2 files changed, 205 insertions(+), 17 deletions(-) create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckAliasTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java index d3c17d5c61..470d5dca97 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java @@ -27,6 +27,18 @@ final class RuleSelfCheck { private RuleSelfCheck() { } + /** 合成 manifest 每片额外占用的行数（EXTINF + URI），用于预判真实规模。 */ + private static final int LINES_PER_SEGMENT_SYNTHETIC = 2; + /** + * 真实 manifest 每片的行数上界。除 EXTINF + URI 外，常见还带 + * {@code #EXT-X-KEY} 或 {@code #EXT-X-PROGRAM-DATE-TIME}。 + * 合成时按这个密度预留余量，否则真实 manifest 会因超过 + * {@code MAX_MANIFEST_LINES} 而回退，而自检看不到。 + */ + private static final int LINES_PER_SEGMENT_REAL = 3; + /** 与 {@code HlsManifestCleaner.MAX_MANIFEST_LINES} 一致。 */ + private static final int MAX_MANIFEST_LINES = 20_000; + /** * @return 载荷是否安全可落地 */ @@ -35,31 +47,58 @@ final class RuleSelfCheck { if (evidence.playlistHost().isEmpty()) return false; if (evidence.inside().isEmpty() || evidence.outside().isEmpty()) return false; + // 真实 manifest 的行密度高于合成，按上界预判：若真实会因规模回退， + // 这条规则在播放时不会生效，不该保存。 + int total = evidence.inside().size() + evidence.outside().size(); + if ((long) total * LINES_PER_SEGMENT_REAL > MAX_MANIFEST_LINES) return false; + HlsManifestCleaner.Rule rule = compile(payload); if (rule == null) return false; + List ordered = ordered(evidence); String base = SYNTHETIC_BASE_SCHEME + evidence.playlistHost() + "/index.m3u8"; - String manifest = synthesize(evidence); HlsManifestCleaner.Result result; try { - result = HlsManifestCleaner.clean(base, manifest, List.of(rule)); + result = HlsManifestCleaner.clean(base, synthesize(evidence, ordered), List.of(rule)); } catch (RuntimeException e) { return false; } // 回退意味着这条规则在真实播放里同样会被拒（全删 / 比例超限 / 时长超限） if (result == null || result.fallback() || !result.changed()) return false; - // 必须恰好删掉区间内的每一片，且一片区间外的都不碰 - return result.removedSegments() == evidence.inside().size() - && !removedAnyOutside(result.manifest(), evidence); + + // 按 URI 出现次数比对，而不是 contains 子串查找 —— 后者会被别名遮蔽： + // 删掉 /s/1 后保留的 /s/12 让判断恒为真。代理式内嵌 URL 的 path、 + // 无扩展名序号切片、同一 URI 在 playlist 中复用都会触发这种遮蔽。 + java.util.Map keptCounts = segmentUriCounts(result.manifest()); + java.util.Map expected = new java.util.HashMap<>(); + for (SegmentFact fact : evidence.outside()) { + expected.merge(uriOf(fact, evidence.playlistHost()), 1, Integer::sum); + } + // 区间外每片都必须保留，且次数完全一致：少一次就是误删。 + if (!expected.equals(keptCounts)) return false; + // 区间内每片都必须被删：与区间外同 URI 的切片会被一并删除， + // 此时上面的次数比对已经失败，这里再兜一次总数。 + return result.removedSegments() == evidence.inside().size(); } - /** 区间外任何一片仍必须留在净化后的 manifest 里。 */ - private static boolean removedAnyOutside(String cleaned, AdIntervalEvidence evidence) { - if (cleaned == null) return true; - for (SegmentFact fact : evidence.outside()) { - if (!cleaned.contains(uriOf(fact, evidence.playlistHost()))) return true; + /** 统计净化后 manifest 里每个切片 URI 的出现次数。 */ + private static java.util.Map segmentUriCounts(String cleaned) { + java.util.Map counts = new java.util.HashMap<>(); + if (cleaned == null) return counts; + for (String line : cleaned.split("\n", -1)) { + String trimmed = line.trim(); + if (trimmed.isEmpty() || trimmed.startsWith("#")) continue; + counts.merge(trimmed, 1, Integer::sum); } - return false; + return counts; + } + + /** 切片按原下标排序，作为合成与回读的共同顺序基准。 */ + private static List ordered(AdIntervalEvidence evidence) { + List all = new java.util.ArrayList<>(evidence.inside()); + all.addAll(evidence.outside()); + all.sort(java.util.Comparator.comparingInt(SegmentFact::index)); + return all; } /** @@ -67,14 +106,13 @@ final class RuleSelfCheck { * *

切片按原下标排序，保留 host、path、时长与断点标记 —— 这些正是规则可能 * 用到的全部条件维度。 + * + *

刻意不注入自定义标签做标记：{@code HlsManifestCleaner} 对未知 tag + * 会整份回退（实测 {@code fallback=true}），注入哨兵反而让自检永远拒绝。 */ - private static String synthesize(AdIntervalEvidence evidence) { - List all = new java.util.ArrayList<>(evidence.inside()); - all.addAll(evidence.outside()); - all.sort(java.util.Comparator.comparingInt(SegmentFact::index)); - + private static String synthesize(AdIntervalEvidence evidence, List ordered) { StringBuilder text = new StringBuilder("#EXTM3U\n#EXT-X-TARGETDURATION:10\n"); - for (SegmentFact fact : all) { + for (SegmentFact fact : ordered) { if (fact.discontinuityBefore()) text.append("#EXT-X-DISCONTINUITY\n"); text.append(String.format(Locale.US, "#EXTINF:%.3f,%n", fact.durationSec())); text.append(uriOf(fact, evidence.playlistHost())).append('\n'); diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckAliasTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckAliasTest.java new file mode 100644 index 0000000000..0c72dfdbb1 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckAliasTest.java @@ -0,0 +1,150 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +/** + * 第六轮评审用随机对抗找出的假阳性：{@code cleaned.contains(uri)} 是子串查找， + * 被删片的 URI 若是某个保留片 URI 的子串，判据就恒为真。 + * + *

配合 {@code removedSegments() == inside.size()} 只比总数， + * 「漏删一片区间内 + 误删一片区间外」会互相抵消而放行。 + */ +public class RuleSelfCheckAliasTest { + + private static final String PLAYLIST_HOST = "v.example.com"; + + private static AdIntervalEvidence evidence(List inside, List outside) { + return new AdIntervalEvidence( + 160_000, 190_000, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + PLAYLIST_HOST, "/play/index.m3u8", true, + inside, outside, true, false, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + /** + * 无扩展名的递增序号切片：删掉 {@code /s/1} 后，保留的 {@code /s/12} + * 让子串判据看不见这次误删。载荷同时命中区间外那片（duration + crossDomain）。 + */ + @Test + public void rejectsWhenRemovedOutsideUriIsSubstringOfKeptOne() { + List inside = new ArrayList<>(); + for (int i = 20; i <= 22; i++) { + inside.add(new SegmentFact(i, "ad.other.com", "/ads/" + (i - 19), 6.4, i == 20)); + } + // 用户多框了一片同域正片，制造「漏删」以抵消下面的「误删」 + inside.add(new SegmentFact(23, PLAYLIST_HOST, "/s/23", 8.0, false)); + + List outside = new ArrayList<>(); + for (int i = 0; i < 20; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/s/" + (100 + i), 8.0, false)); + } + // 会被 duration+crossDomain 命中的正片，其 URI 是下一片的前缀 + outside.add(new SegmentFact(24, "cdn2.net", "/s/1", 6.4, false)); + outside.add(new SegmentFact(25, "cdn2.net", "/s/12", 8.0, false)); + + RulePayload payload = RulePayload.ofHlsRule( + List.of(PLAYLIST_HOST), List.of("ad.other.com"), List.of(), + 6.3, 6.5, false, true, 2); + + assertFalse("区间外正片被误删时必须拒绝，不能因子串遮蔽而放行", + RuleSelfCheck.isSafe(evidence(inside, outside), payload)); + } + + /** 代理式内嵌 URL 的 path 会原样保留，同样能遮蔽真实删除。 */ + @Test + public void rejectsWhenProxyStylePathShadowsRemovedSegment() { + List inside = List.of( + new SegmentFact(2, "ad.other.com", "/seg/1.ts", 8.0, true), + new SegmentFact(3, "edge7." + PLAYLIST_HOST, "/ads/1.ts", 8.0, false)); + List outside = List.of( + new SegmentFact(0, "ad.other.com", "/a.ts", 8.0, false), + new SegmentFact(1, "edge7." + PLAYLIST_HOST, "/x/ads/1.ts", 6.4, false), + // 内嵌完整 URL：pathOf 会原样返回，成为上面 index=0 的遮蔽者 + new SegmentFact(4, "edge7." + PLAYLIST_HOST, + "/p/https://ad.other.com/a.ts", 10.0, false), + new SegmentFact(5, "shared-cdn.net", "/seg/1.ts", 10.0, false)); + + RulePayload payload = RulePayload.ofHlsRule( + List.of(PLAYLIST_HOST), List.of("ad.other.com"), List.of(), + Double.NaN, Double.NaN, false, true, 2); + + assertFalse("内嵌 URL 的 path 不得遮蔽区间外的误删", + RuleSelfCheck.isSafe(evidence(inside, outside), payload)); + } + + /** 同一 URI 在 playlist 中复用（片头片尾同一 promo）不得造成误判。 */ + @Test + public void rejectsWhenDuplicateUriAppearsInsideAndOutside() { + List inside = List.of( + new SegmentFact(10, "ad.other.com", "/promo.ts", 6.4, true), + new SegmentFact(11, "ad.other.com", "/promo.ts", 6.4, false)); + List outside = new ArrayList<>(); + for (int i = 0; i < 10; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/s/" + i + ".ts", 8.0, false)); + } + // 片尾复用同一个 promo 切片：它会被同一条规则删掉 + outside.add(new SegmentFact(29, "ad.other.com", "/promo.ts", 6.4, false)); + for (int i = 12; i < 29; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/s/" + i + ".ts", 8.0, i == 12)); + } + + RulePayload payload = RulePayload.ofHlsRule( + List.of(PLAYLIST_HOST), List.of("ad.other.com"), List.of(), + Double.NaN, Double.NaN, false, true, 2); + + assertFalse("区间外同 URI 的切片会被一并删除，必须拒绝", + RuleSelfCheck.isSafe(evidence(inside, outside), payload)); + } + + /** 规模预判：真实 manifest 行密度更高，接近上限时必须提前拒绝。 */ + @Test + public void rejectsWhenRealManifestWouldExceedLineLimit() + { + List inside = new ArrayList<>(); + for (int i = 0; i < 3; i++) { + inside.add(new SegmentFact(i, "ad.other.com", "/ads/" + i + ".ts", 6.4, i == 0)); + } + List outside = new ArrayList<>(); + // 合成 2 行/片时不超 20000，真实 3 行/片时会超 + for (int i = 3; i < 7_000; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/s/" + i + ".ts", 8.0, i == 3)); + } + + RulePayload payload = RulePayload.ofHlsRule( + List.of(PLAYLIST_HOST), List.of("ad.other.com"), List.of(), + Double.NaN, Double.NaN, false, true, 2); + + assertFalse("真实 manifest 会因行数超限回退，规则不该保存", + RuleSelfCheck.isSafe(evidence(inside, outside), payload)); + } + + /** 健康场景仍须放行，否则自检等于把功能关掉。 */ + @Test + public void stillAcceptsCleanlySeparableAdBlock() { + List inside = new ArrayList<>(); + for (int i = 10; i < 13; i++) { + inside.add(new SegmentFact(i, "ad-cdn.other.com", "/ads/" + i + ".ts", 6.4, i == 10)); + } + List outside = new ArrayList<>(); + for (int i = 0; i < 10; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/seg/" + i + ".ts", 8.0, false)); + } + for (int i = 13; i < 30; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/seg/" + i + ".ts", 8.0, i == 13)); + } + + RulePayload payload = RulePayload.ofHlsRule( + List.of(PLAYLIST_HOST), List.of("ad-cdn.other.com"), List.of(), + Double.NaN, Double.NaN, false, true, 2); + + assertTrue("干净可分的广告块必须放行", + RuleSelfCheck.isSafe(evidence(inside, outside), payload)); + } +} From 3ff99a8cc7ba0b9d151cdf2e790c9f094c588205 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 13:41:30 +0800 Subject: [PATCH 25/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E8=B7=AF?= =?UTF-8?q?=E5=BE=84=E6=AD=A3=E5=88=99=E8=B7=A8=20fragment=20=E8=AF=AF?= =?UTF-8?q?=E5=88=A0=E6=AD=A3=E7=89=87=EF=BC=8C=E5=90=AF=E7=94=A8=E5=B7=B2?= =?UTF-8?q?=E6=9C=89=E8=A7=84=E5=88=99=E8=A1=A5=E4=B8=8A=E8=87=AA=E6=A3=80?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 第七轮评审用 30 万次随机对抗给出分层结论：RuleSelfCheck 作为独立 API 仍有 3% 假阳性（全部是同 URI 一删一留的计数抵消），但两个分类器实际能产出的载荷 形状下闭合（0/7571）—— 因为 payloadOf 恒置 requireDiscontinuity=false， 同 URI 又必然让路径条件对两侧同时成立，凑不够 minimumSignals。第六轮的 contains 子串遮蔽已确认修掉。 本次修三个端到端可达的问题： 1. pathOnlyPattern 的 ^[^?]* 只挡 query，不挡 fragment。证据侧用 URI.getPath()，fragment 被丢弃，所以 pathAbsentOutside 与 RuleSelfCheck 都看不见它。实测正片 URL 形如 /seg/99.ts#/ads/x 会被跨 # 命中， 真实 cleaner removedSegments=3（应为 2）且 fallback=false —— 错误不被回退 兜住。在专门生成该形状的 20 万轮里，放行载荷 87.7% 会误删正片。 改为 ^[^?#]*。 2. ENABLE_EXISTING_RULE 完全绕过 RuleSelfCheck，而它的 score 0.825 高于 HLS 规则通道，30 万轮里被选为首选 63984 次。判据只看「hostSuffixes 与 区间内某片同域」，不看该规则的其余条件：一条 minimumSignals=1 的既有规则 在「广告与正片共用同一 CDN」时命中全部切片，回退并连带停掉 legacy。 RuleState 新增 compiled 字段携带编译产物，findDisabledMatch 用 RuleSelfCheck.isSafe 实跑验证；拿不到编译产物时不建议启用。 3. 行密度系数 3 双向都不准。真实 KEY+PDT+DISCONTINUITY 是 5 行/片且实测可 正常净化，系数 3 会让 5000..6666 片的 playlist 误放行 —— 规则存下来却因 行数超限永久无效。提到 5：误拒方向落在 6667..9998 片（4s/片约 7.4 小时）， 正片剧集触达不到，代价可忽略；误放行才是实际损失。 原有测试的 RuleState fixture 都用简化构造器（compiled=null）且区间外只给 一两片对照，删除比例天然超 35%，新逻辑正确地拒绝了它们，一并补足。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback 144 tests 0 failures; full suite 4041 tests with the same 5 pre-existing failures as baseline; both flavors compile Task-Guard: AF5l-review7 --- .../tv/ad/feedback/AdFeedbackDataSource.java | 22 ++++++- .../ad/feedback/ExistingRuleClassifier.java | 26 +++++++- .../tv/ad/feedback/HlsSegmentClassifier.java | 6 +- .../android/tv/ad/feedback/RuleSelfCheck.java | 31 +++++++--- .../ad/feedback/AdFeedbackControllerTest.java | 3 +- .../ad/feedback/AdFeedbackDataSourceTest.java | 12 +++- .../feedback/ExistingRuleClassifierTest.java | 62 ++++++++++++++++++- .../tv/ad/feedback/PathFragmentTest.java | 37 +++++++++++ .../tv/ad/feedback/RulePayloadTest.java | 5 +- 9 files changed, 186 insertions(+), 18 deletions(-) create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/PathFragmentTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java index 15d864cb42..3526faeb83 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java @@ -1,11 +1,14 @@ package com.fongmi.android.tv.ad.feedback; +import com.fongmi.android.tv.App; import com.fongmi.android.tv.api.config.HlsRuleConfig; import com.fongmi.android.tv.api.config.ImportedAdRuleCandidateStore; import com.fongmi.android.tv.api.config.RuleConfig; import com.fongmi.android.tv.api.config.UserAdRuleStore; +import com.fongmi.android.tv.bean.HlsAdRule; import com.fongmi.android.tv.bean.ImportedAdRuleCandidate; import com.fongmi.android.tv.bean.UserAdRule; +import com.fongmi.android.tv.utils.HlsManifestCleaner; import com.google.gson.JsonArray; import com.google.gson.JsonElement; import com.google.gson.JsonObject; @@ -93,7 +96,8 @@ public final class AdFeedbackDataSource { states.add(new ExistingRuleClassifier.RuleState( entry.key(), entry.id(), entry.name(), entry.enabled(), entry.valid(), - hostSuffixesOf(entry.detail()))); + hostSuffixesOf(entry.detail()), + compiledOf(entry.detail()))); } return List.copyOf(states); } catch (RuntimeException e) { @@ -101,6 +105,22 @@ public final class AdFeedbackDataSource { } } + /** + * 从规则详情 JSON 还原已编译的规则，供 {@code RuleSelfCheck} 实跑验证。 + * + *

{@code HlsRuleConfig.Entry} 只暴露 detail JSON，不给编译产物；这里重新 + * 反序列化并 compile。失败返回 null，调用方会因此不建议启用该规则。 + */ + static HlsManifestCleaner.Rule compiledOf(String detailJson) { + if (detailJson == null || detailJson.isBlank()) return null; + try { + HlsAdRule rule = App.gson().fromJson(detailJson, HlsAdRule.class); + return rule == null ? null : rule.compile(); + } catch (RuntimeException e) { + return null; + } + } + /** 从规则详情 JSON 里解析 hostSuffixes。解析失败返回空列表，不抛异常。 */ static List hostSuffixesOf(String detailJson) { if (detailJson == null || detailJson.isBlank()) return List.of(); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java index 863dc6514e..a468a9e8c0 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java @@ -1,5 +1,7 @@ package com.fongmi.android.tv.ad.feedback; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + import java.util.ArrayList; import java.util.List; import java.util.Locale; @@ -29,9 +31,11 @@ public final class ExistingRuleClassifier { * @param enabled 当前是否启用 * @param valid 编译是否通过 * @param hostSuffixes 该规则针对的切片域名后缀 + * @param compiled 已编译的规则，供自检实跑；取不到时为 null（则不建议启用） */ public record RuleState(String key, String id, String name, boolean enabled, - boolean valid, List hostSuffixes) { + boolean valid, List hostSuffixes, + HlsManifestCleaner.Rule compiled) { public RuleState { key = key == null ? "" : key; @@ -39,6 +43,12 @@ public final class ExistingRuleClassifier { name = name == null ? "" : name; hostSuffixes = hostSuffixes == null ? List.of() : List.copyOf(hostSuffixes); } + + /** 无编译产物的简化构造，仅用于诊断路径与测试。 */ + public RuleState(String key, String id, String name, boolean enabled, + boolean valid, List hostSuffixes) { + this(key, id, name, enabled, valid, hostSuffixes, null); + } } /** @@ -102,12 +112,24 @@ public final class ExistingRuleClassifier { } /** 找出一条覆盖本区间域名、有效但未启用的规则。 */ + /** + * 找出一条覆盖本区间域名、有效但未启用的规则。 + * + *

必须过 {@link RuleSelfCheck}：仅凭「hostSuffixes 与区间内某片同域」建议启用， + * 不看该规则的其余条件 —— 实测一条 {@code minimumSignals=1} 的既有规则在 + * 「广告与正片共用同一 CDN」时会命中全部切片，回退并连带停掉 legacy 启发式。 + * 这条路径此前完全没有守卫。 + */ private static RuleState findDisabledMatch(AdIntervalEvidence evidence, List rules) { for (RuleState rule : rules) { if (rule.enabled() || !rule.valid() || rule.hostSuffixes().isEmpty()) continue; boolean covers = evidence.inside().stream() .anyMatch(fact -> rule.hostSuffixes().stream().anyMatch(fact::hostEndsWith)); - if (covers) return rule; + if (!covers) continue; + // 拿不到编译产物时无法验证，宁可不建议 + if (rule.compiled() == null) continue; + if (!RuleSelfCheck.isSafe(evidence, rule.compiled())) continue; + return rule; } return null; } diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index 4eb84a9d4c..d3a561d536 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -194,7 +194,11 @@ public final class HlsSegmentClassifier { * legacy 启发式。{@code [^?]*} 无法跨过 {@code ?}，从而把匹配限制在 path 内。 */ static String pathOnlyPattern(String hint) { - return "^[^?]*" + Pattern.quote(hint); + // 同时排除 query 与 fragment：证据侧用 URI.getPath()，query 和 fragment + // 都被丢弃，所以 pathAbsentOutside 与 RuleSelfCheck 都看不见它们。 + // 只挡 ? 的话，正片 URL 形如 /seg/99.ts#/ads/x 会被跨 # 命中 —— 实测 + // 这种形状下真实 cleaner 会多删一片正片且 fallback=false，错误不被兜住。 + return "^[^?#]*" + Pattern.quote(hint); } /** host 是否等于给定域名或为其子域。 */ diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java index 470d5dca97..248d01e157 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java @@ -31,11 +31,17 @@ final class RuleSelfCheck { private static final int LINES_PER_SEGMENT_SYNTHETIC = 2; /** * 真实 manifest 每片的行数上界。除 EXTINF + URI 外，常见还带 - * {@code #EXT-X-KEY} 或 {@code #EXT-X-PROGRAM-DATE-TIME}。 - * 合成时按这个密度预留余量，否则真实 manifest 会因超过 - * {@code MAX_MANIFEST_LINES} 而回退，而自检看不到。 + * {@code #EXT-X-KEY}、{@code #EXT-X-PROGRAM-DATE-TIME} 与 + * {@code #EXT-X-DISCONTINUITY}，实测这种 5 行/片的 playlist 能正常净化。 + * + *

取上界而非典型值：低估会让规则在真实 manifest 上因行数超限而回退 —— + * 规则存下来却永久无效，用户看到「已保存」但广告依旧。高估只是让超长 + * playlist 弃权，而按 4s/片算 4000 片已是 4.4 小时，正片剧集触达不到。 + * + *

{@code #EXT-X-BYTERANGE} 不计入：{@code HlsManifestCleaner} 对含该标签的 + * manifest 无条件回退，这类 playlist 从不参与净化。 */ - private static final int LINES_PER_SEGMENT_REAL = 3; + private static final int LINES_PER_SEGMENT_REAL = 5; /** 与 {@code HlsManifestCleaner.MAX_MANIFEST_LINES} 一致。 */ private static final int MAX_MANIFEST_LINES = 20_000; @@ -44,6 +50,20 @@ final class RuleSelfCheck { */ static boolean isSafe(AdIntervalEvidence evidence, RulePayload payload) { if (payload.isEmpty()) return false; + HlsManifestCleaner.Rule rule = compile(payload); + if (rule == null) return false; + return isSafe(evidence, rule); + } + + /** + * 直接校验一条已编译的规则，供「启用已有规则」路径复用。 + * + *

那条路径此前完全绕过自检：它只判断规则的 hostSuffixes 与区间内某片同域， + * 不看该规则的其余条件。实测一条 {@code minimumSignals=1} 的既有规则在 + * 「广告与正片共用同一 CDN」时会命中全部切片。 + */ + static boolean isSafe(AdIntervalEvidence evidence, HlsManifestCleaner.Rule rule) { + if (rule == null) return false; if (evidence.playlistHost().isEmpty()) return false; if (evidence.inside().isEmpty() || evidence.outside().isEmpty()) return false; @@ -52,9 +72,6 @@ final class RuleSelfCheck { int total = evidence.inside().size() + evidence.outside().size(); if ((long) total * LINES_PER_SEGMENT_REAL > MAX_MANIFEST_LINES) return false; - HlsManifestCleaner.Rule rule = compile(payload); - if (rule == null) return false; - List ordered = ordered(evidence); String base = SYNTHETIC_BASE_SCHEME + evidence.playlistHost() + "/index.m3u8"; HlsManifestCleaner.Result result; diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java index df903f9064..bd44351bda 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java @@ -214,7 +214,8 @@ public class AdFeedbackControllerTest { FakeHost host = new FakeHost(); host.playlist = adBlockPlaylist(); host.ruleStates = List.of(new ExistingRuleClassifier.RuleState( - "builtin:x", "x-rule", "实验规则", false, true, List.of("ad-cdn.other.com"))); + "builtin:x", "x-rule", "实验规则", false, true, List.of("ad-cdn.other.com"), + ExistingRuleClassifierTest.compiledRule("ad-cdn.other.com"))); AdFeedbackController controller = new AdFeedbackController(host); controller.onMarkedInterval(30_000, 49_200); diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSourceTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSourceTest.java index 7e750d33f3..0eadbd2275 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSourceTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSourceTest.java @@ -58,14 +58,22 @@ public class AdFeedbackDataSourceTest { List suffixes = AdFeedbackDataSource.hostSuffixesOf( "{\"hostSuffixes\":[\"ad-cdn.other.com\"]}"); ExistingRuleClassifier.RuleState disabled = new ExistingRuleClassifier.RuleState( - "builtin:x", "x", "实验规则", false, true, suffixes); + "builtin:x", "x", "实验规则", false, true, suffixes, + ExistingRuleClassifierTest.compiledRule("ad-cdn.other.com")); + List outside = new java.util.ArrayList<>(); + for (int i = 0; i < 3; i++) { + outside.add(new SegmentFact(i, "v.example.com", "/seg/" + i + ".ts", 8.0, false)); + } + for (int i = 4; i < 30; i++) { + outside.add(new SegmentFact(i, "v.example.com", "/seg/" + i + ".ts", 8.0, i == 4)); + } AdIntervalEvidence evidence = new AdIntervalEvidence( 10_000, 40_000, StartOrigin.USER_MARKED, "site", "站点", "剧名", "线路", "第 1 集", "v.example.com", "/play/index.m3u8", true, List.of(new SegmentFact(3, "ad-cdn.other.com", "/seg/3.ts", 6.4, true)), - List.of(new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false)), + outside, false, false, List.of(), false, List.of(), AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifierTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifierTest.java index bee31c8d19..04eef1eb52 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifierTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifierTest.java @@ -13,22 +13,40 @@ public class ExistingRuleClassifierTest { private static final String PLAYLIST_HOST = "v.example.com"; + /** + * 区间外放足量同域正片：cleaner 有 35% 删除比例上限，只给一两片对照时 + * 任何规则都会因超限回退，「启用已有规则」的自检会正确拒掉。 + */ private static AdIntervalEvidence evidence(String insideHost, List removedRuleIds) { + List outside = new java.util.ArrayList<>(); + for (int i = 0; i < 3; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/seg/" + i + ".ts", 8.0, false)); + } + for (int i = 4; i < 30; i++) { + outside.add(new SegmentFact(i, PLAYLIST_HOST, "/seg/" + i + ".ts", 8.0, i == 4)); + } return new AdIntervalEvidence( 10_000, 40_000, StartOrigin.USER_MARKED, "site", "站点", "剧名", "线路", "第 1 集", PLAYLIST_HOST, "/play/index.m3u8", true, List.of(new SegmentFact(3, insideHost, "/seg/3.ts", 6.4, true)), - List.of(new SegmentFact(0, PLAYLIST_HOST, "/seg/0.ts", 8.0, false)), + outside, false, false, removedRuleIds, false, List.of(), AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); } + /** 一条只命中给定切片域名、带作用域的可编译规则。 */ + static com.fongmi.android.tv.utils.HlsManifestCleaner.Rule compiledRule(String hostSuffix) { + return com.fongmi.android.tv.bean.HlsAdRule.createUserRule( + "id", "n", List.of(PLAYLIST_HOST), List.of(hostSuffix), List.of(), + null, null, false, true, 2).compile(); + } + @Test public void suggestsEnablingDisabledRuleInsteadOfCreatingNew() { ExistingRuleClassifier.RuleState disabled = new ExistingRuleClassifier.RuleState( "builtin:baofeng", "baofeng-preroll", "暴风片头", false, true, - List.of("ad-cdn.other.com")); + List.of("ad-cdn.other.com"), compiledRule("ad-cdn.other.com")); AdAttribution attribution = ExistingRuleClassifier.classify( evidence("ad-cdn.other.com", List.of()), @@ -109,4 +127,44 @@ public class ExistingRuleClassifierTest { ExistingRuleClassifier.Input.empty())); assertNull(ExistingRuleClassifier.classify(null, ExistingRuleClassifier.Input.empty())); } + + /** + * 既有规则过宽时不得建议启用。 + * + *

此前这条路径完全没有守卫 —— 只判断 hostSuffixes 与区间内某片同域， + * 不看规则的其余条件。一条 {@code minimumSignals=1} 且只限定 playlist 作用域的 + * 规则会命中该站每一片切片，启用后 cleaner 因全删而回退，并连带停掉 legacy。 + */ + @Test + public void doesNotSuggestEnablingAnOverbroadRule() { + com.fongmi.android.tv.utils.HlsManifestCleaner.Rule overbroad = + com.fongmi.android.tv.bean.HlsAdRule.createUserRule( + "wide", "wide", List.of(PLAYLIST_HOST), List.of("ad-cdn.other.com"), + List.of(), 0.0, 3600.0, false, false, 1).compile(); + ExistingRuleClassifier.RuleState wide = new ExistingRuleClassifier.RuleState( + "builtin:wide", "wide", "过宽规则", false, true, + List.of("ad-cdn.other.com"), overbroad); + + AdAttribution attribution = ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of()), + new ExistingRuleClassifier.Input(List.of(wide), false, List.of())); + + // 要么整体弃权，要么只报诊断，绝不能给出 ENABLE_EXISTING_RULE + assertTrue(attribution == null + || attribution.remediation() != RemediationKind.ENABLE_EXISTING_RULE); + } + + /** 拿不到编译产物时无法验证，宁可不建议启用。 */ + @Test + public void doesNotSuggestEnablingWhenCompiledRuleIsUnavailable() { + ExistingRuleClassifier.RuleState noCompiled = new ExistingRuleClassifier.RuleState( + "builtin:x", "x", "无编译产物", false, true, List.of("ad-cdn.other.com")); + + AdAttribution attribution = ExistingRuleClassifier.classify( + evidence("ad-cdn.other.com", List.of()), + new ExistingRuleClassifier.Input(List.of(noCompiled), false, List.of())); + + assertTrue(attribution == null + || attribution.remediation() != RemediationKind.ENABLE_EXISTING_RULE); + } } diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/PathFragmentTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/PathFragmentTest.java new file mode 100644 index 0000000000..fee28d2987 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/PathFragmentTest.java @@ -0,0 +1,37 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import org.junit.Test; + +import java.util.regex.Pattern; + +/** + * 路径正则必须同时排除 query 与 fragment。 + * + *

证据侧用 {@code URI.getPath()}，query 和 fragment 都被丢弃，所以 + * {@code SegmentContrast.pathAbsentOutside} 与 {@code RuleSelfCheck} 都看不见它们。 + * 只挡 {@code ?} 时，正片 URL 形如 {@code /seg/99.ts#/ads/x} 会被跨 {@code #} 命中， + * 实测真实 cleaner 会多删一片正片且 {@code fallback=false}，错误不被兜住。 + */ +public class PathFragmentTest { + + @Test + public void patternRejectsBothQueryAndFragment() { + Pattern p = Pattern.compile(HlsSegmentClassifier.pathOnlyPattern("/ads/")); + + assertTrue("真实广告目录段必须命中", p.matcher("https://v.example.com/ads/1.ts").find()); + + assertFalse("query 里的 hint 不得命中", + p.matcher("https://v.example.com/seg/1.ts?ref=/ads/").find()); + assertFalse("fragment 里的 hint 不得命中", + p.matcher("https://v.example.com/seg/99.ts#/ads/x").find()); + assertFalse("query + fragment 组合同样不得命中", + p.matcher("https://v.example.com/seg/9.ts?t=1#/ads/x").find()); + assertFalse("域名含 ads 不得命中", + p.matcher("https://ads.example.com/seg/1.ts").find()); + assertFalse("无斜杠包围不得命中", + p.matcher("https://v.example.com/upload/ads1.ts").find()); + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java index c25f953b58..ea68f05aee 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RulePayloadTest.java @@ -178,13 +178,14 @@ public class RulePayloadTest { public void existingRuleAttributionCarriesRuleKey() { ExistingRuleClassifier.RuleState disabled = new ExistingRuleClassifier.RuleState( "builtin|pkg|baofeng", "baofeng", "暴风片头", false, true, - List.of("ad-cdn.other.com")); + List.of("ad-cdn.other.com"), + ExistingRuleClassifierTest.compiledRule("ad-cdn.other.com")); AdIntervalEvidence evidence = new AdIntervalEvidence( 10_000, 40_000, StartOrigin.USER_MARKED, "site", "站点", "剧名", "线路", "第 1 集", "v.example.com", "/play/index.m3u8", true, List.of(new SegmentFact(3, "ad-cdn.other.com", "/seg/3.ts", 6.4, true)), - List.of(new SegmentFact(0, "v.example.com", "/seg/0.ts", 8.0, false)), + healthyOutside(3, 1), false, false, List.of(), false, List.of(), AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); From c5acc64d9a3188b056bb58d901cdaedfe6e3e2fd Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 15:27:52 +0800 Subject: [PATCH 26/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E6=97=A2?= =?UTF-8?q?=E6=9C=89=E8=A7=84=E5=88=99=E5=B8=A6=E6=9C=AA=E9=94=9A=E5=AE=9A?= =?UTF-8?q?=E5=88=87=E7=89=87=E6=AD=A3=E5=88=99=E6=97=B6=E4=BC=9A=E9=9D=99?= =?UTF-8?q?=E9=BB=98=E8=AF=AF=E5=88=A0=E6=AD=A3=E7=89=87?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 第八轮评审复现出上一轮修复未覆盖的链路：RuleSelfCheck 的合成 manifest 用 SegmentFact.path() 拼 URL，而那是 URI.getPath() 的产物 —— 按去敏要求 query 与 fragment 都被丢弃。于是合成 URL 永不带 ? 和 #，而真实 matchesPattern 匹配的是 含两者的完整 URL。 上一轮给自己生成的正则加 ^[^?#]* 前缀对齐了这个差异，但既有规则的正则是第三方 写的，不会有这个锚定。实测一条 segmentUrlRegex=["/ads/"] 的既有规则： 自检放行，真实 cleaner removed=5（应为 3），超删 5/40=12.5% 低于 35% 闸门故 fallback=false —— 错误不被回退兜住，用户直接看到正片跳帧。 这个场景的现实性不低：HlsRuleConfig.compileExternal 对外部规则用 Boolean.TRUE.equals(enabled) 判定启用状态，即接口下发的 hlsRules 只要没显式写 "enabled":true 就落地为 disabled+valid，正是「启用已有规则」通道的主要来源。 - compiledOf 新增 hasUnanchoredSegmentRegex 前置检查：凡带 segmentUrlRegex 且 未以 ^[^?#]* 开头的规则一律返回 null，不为它背书。解析失败也保守判为未锚定。 - compiledOf 从 App.gson() 改为本地 Gson：前者在纯 JVM 单测里 App.get() 为 null 会抛 NPE，被 catch 吞掉后静默让整个通道弃权，这条路径因此既没测也测不了。 评审实测确认两者的 @SerializedName 映射与 disableHtmlEscaping 往返完全等价。 - 行数预判补 16 行 header 余量：total*5 在 total=4000 时放行，但真实 manifest 的 #EXTM3U / #EXT-X-VERSION / #EXT-X-ENDLIST 等会让 lineCount 越过 20000 而回退。 - RuleState 的 6 参简化构造器从 public 收为包级：生产误用会静默关掉整条通道。 - 清掉重复 Javadoc 与死常量 LINES_PER_SEGMENT_SYNTHETIC。 评审确认无回归：全仓 7 个 RuleState 构造点，生产侧唯一一处已传 compiledOf， 测试侧 3 处用简化构造器的都是纯诊断断言。三处修复逐项回退实测均有测试捕获。 未修（记录在案）：自检只跑 List.of(rule)，不校验与站内已启用规则叠加后是否 越过 35% 闸门。这是 HLS_STRUCTURED_RULE 路径的既有局限，需要给 RuleSelfCheck 传入当前已启用规则列表，改动面较大，留待后续。 Verification: testMobileArm64_v8aDebugUnitTest ad.feedback 150 tests 0 failures (CompiledRuleGateTest 6 new); full suite 4047 tests with the same 5 pre-existing failures as baseline; both flavors compile Task-Guard: AF5m-review8 --- .../tv/ad/feedback/AdFeedbackDataSource.java | 58 ++++++- .../ad/feedback/ExistingRuleClassifier.java | 13 +- .../android/tv/ad/feedback/RuleSelfCheck.java | 12 +- .../tv/ad/feedback/CompiledRuleGateTest.java | 149 ++++++++++++++++++ 4 files changed, 222 insertions(+), 10 deletions(-) create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/CompiledRuleGateTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java index 3526faeb83..f88cc37f6f 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java @@ -1,6 +1,5 @@ package com.fongmi.android.tv.ad.feedback; -import com.fongmi.android.tv.App; import com.fongmi.android.tv.api.config.HlsRuleConfig; import com.fongmi.android.tv.api.config.ImportedAdRuleCandidateStore; import com.fongmi.android.tv.api.config.RuleConfig; @@ -9,6 +8,7 @@ import com.fongmi.android.tv.bean.HlsAdRule; import com.fongmi.android.tv.bean.ImportedAdRuleCandidate; import com.fongmi.android.tv.bean.UserAdRule; import com.fongmi.android.tv.utils.HlsManifestCleaner; +import com.google.gson.Gson; import com.google.gson.JsonArray; import com.google.gson.JsonElement; import com.google.gson.JsonObject; @@ -27,6 +27,15 @@ import java.util.Set; */ public final class AdFeedbackDataSource { + /** + * 与 {@code HlsSegmentClassifier.pathOnlyPattern} 生成的前缀一致。 + * 只有以它开头的切片正则，其匹配范围才与自检的合成 manifest 语义相同。 + */ + static final String ANCHORED_PATH_PREFIX = "^[^?#]*"; + + /** 本地 Gson，避免依赖 Application 生命周期，使 compiledOf 可被单测。 */ + private static final Gson GSON = new Gson(); + private AdFeedbackDataSource() { } @@ -110,17 +119,62 @@ public final class AdFeedbackDataSource { * *

{@code HlsRuleConfig.Entry} 只暴露 detail JSON，不给编译产物；这里重新 * 反序列化并 compile。失败返回 null，调用方会因此不建议启用该规则。 + * + *

用本地 Gson 而非 {@code App.gson()}：后者在纯 JVM 单测里 + * {@code App.get()} 为 null 会抛 NPE，被 catch 吞掉后静默让整个通道弃权， + * 这条路径因此既没测也测不了。 + * + *

带未锚定 {@code segmentUrlRegex} 的规则一律拒绝，见 + * {@link #hasUnanchoredSegmentRegex}。 */ static HlsManifestCleaner.Rule compiledOf(String detailJson) { if (detailJson == null || detailJson.isBlank()) return null; + if (hasUnanchoredSegmentRegex(detailJson)) return null; try { - HlsAdRule rule = App.gson().fromJson(detailJson, HlsAdRule.class); + HlsAdRule rule = GSON.fromJson(detailJson, HlsAdRule.class); return rule == null ? null : rule.compile(); } catch (RuntimeException e) { return null; } } + /** + * 规则的切片正则是否未锚定到 path。 + * + *

{@code RuleSelfCheck} 的合成 manifest 用 {@code SegmentFact.path()} 拼 URL， + * 而那是 {@code URI.getPath()} 的产物 —— 按去敏要求 query 与 fragment 都被丢弃。 + * 于是合成 URL 永不带 {@code ?}/{@code #}，而真实 {@code matchesPattern} 匹配的是 + * 含 query 与 fragment 的完整 URL。 + * + *

本项目自己生成的正则用 {@code ^[^?#]*} 前缀对齐了这个差异（见 + * {@code HlsSegmentClassifier.pathOnlyPattern}），但接口下发或内置的第三方规则 + * 不会有这个锚定。实测一条 {@code segmentUrlRegex=["/ads/"]} 的既有规则在 + * 「正片 URL 带 ?ref=/ads/」时自检放行、真实运行多删正片且 {@code fallback=false} + * —— 错误不被回退兜住，用户直接看到跳帧。 + * + *

因此凡是带 {@code segmentUrlRegex} 且未显式排除 query/fragment 的规则， + * 都不能用自检结论为它背书。 + */ + static boolean hasUnanchoredSegmentRegex(String detailJson) { + if (detailJson == null || detailJson.isBlank()) return false; + try { + JsonElement parsed = JsonParser.parseString(detailJson); + if (!parsed.isJsonObject()) return false; + JsonElement regex = parsed.getAsJsonObject().get("segmentUrlRegex"); + if (regex == null || !regex.isJsonArray()) return false; + for (JsonElement element : regex.getAsJsonArray()) { + if (element == null || element.isJsonNull()) continue; + String value = element.getAsString(); + if (value == null || value.isBlank()) continue; + if (!value.startsWith(ANCHORED_PATH_PREFIX)) return true; + } + return false; + } catch (RuntimeException e) { + // 解析不出来时保守判为未锚定 + return true; + } + } + /** 从规则详情 JSON 里解析 hostSuffixes。解析失败返回空列表，不抛异常。 */ static List hostSuffixesOf(String detailJson) { if (detailJson == null || detailJson.isBlank()) return List.of(); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java index a468a9e8c0..d000c518c5 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java @@ -44,9 +44,15 @@ public final class ExistingRuleClassifier { hostSuffixes = hostSuffixes == null ? List.of() : List.copyOf(hostSuffixes); } - /** 无编译产物的简化构造，仅用于诊断路径与测试。 */ - public RuleState(String key, String id, String name, boolean enabled, - boolean valid, List hostSuffixes) { + /** + * 无编译产物的简化构造，仅用于诊断路径与测试。 + * + *

刻意保持包级可见：生产代码必须走 7 参构造并提供编译产物， + * 否则 {@code findDisabledMatch} 会因无法自检而跳过该规则 —— + * 用它会静默关掉「启用已有规则」整条通道。 + */ + RuleState(String key, String id, String name, boolean enabled, + boolean valid, List hostSuffixes) { this(key, id, name, enabled, valid, hostSuffixes, null); } } @@ -111,7 +117,6 @@ public final class ExistingRuleClassifier { 0.5f, RiskLevel.LOW, lines, RemediationKind.NONE); } - /** 找出一条覆盖本区间域名、有效但未启用的规则。 */ /** * 找出一条覆盖本区间域名、有效但未启用的规则。 * diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java index 248d01e157..9213512892 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java @@ -27,8 +27,6 @@ final class RuleSelfCheck { private RuleSelfCheck() { } - /** 合成 manifest 每片额外占用的行数（EXTINF + URI），用于预判真实规模。 */ - private static final int LINES_PER_SEGMENT_SYNTHETIC = 2; /** * 真实 manifest 每片的行数上界。除 EXTINF + URI 外，常见还带 * {@code #EXT-X-KEY}、{@code #EXT-X-PROGRAM-DATE-TIME} 与 @@ -44,6 +42,8 @@ final class RuleSelfCheck { private static final int LINES_PER_SEGMENT_REAL = 5; /** 与 {@code HlsManifestCleaner.MAX_MANIFEST_LINES} 一致。 */ private static final int MAX_MANIFEST_LINES = 20_000; + /** 为真实 manifest 的头部标签预留的行数余量。 */ + private static final int MANIFEST_HEADER_LINES = 16; /** * @return 载荷是否安全可落地 @@ -68,9 +68,13 @@ final class RuleSelfCheck { if (evidence.inside().isEmpty() || evidence.outside().isEmpty()) return false; // 真实 manifest 的行密度高于合成，按上界预判：若真实会因规模回退， - // 这条规则在播放时不会生效，不该保存。 + // 这条规则在播放时不会生效，不该保存。除切片行外还要留出 header + // （#EXTM3U / #EXT-X-VERSION / #EXT-X-TARGETDURATION / + // #EXT-X-MEDIA-SEQUENCE / #EXT-X-ENDLIST 等）的余量，否则恰好卡在 + // 边界的 playlist 会被放行而真实仍然回退。 int total = evidence.inside().size() + evidence.outside().size(); - if ((long) total * LINES_PER_SEGMENT_REAL > MAX_MANIFEST_LINES) return false; + if ((long) total * LINES_PER_SEGMENT_REAL + MANIFEST_HEADER_LINES + > MAX_MANIFEST_LINES) return false; List ordered = ordered(evidence); String base = SYNTHETIC_BASE_SCHEME + evidence.playlistHost() + "/index.m3u8"; diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/CompiledRuleGateTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/CompiledRuleGateTest.java new file mode 100644 index 0000000000..abed517f4c --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/CompiledRuleGateTest.java @@ -0,0 +1,149 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertNotNull; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import com.fongmi.android.tv.bean.HlsAdRule; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +/** + * 「启用已有规则」通道的守卫：{@code compiledOf} 必须拒掉那些自检会放行、 + * 但真实运行会误删正片的既有规则。 + * + *

第八轮评审复现的链路：{@code RuleSelfCheck} 的合成 manifest 用 + * {@code SegmentFact.path()} 拼 URL，而那是 {@code URI.getPath()} 的产物， + * query 与 fragment 都被去敏丢弃。于是一条 {@code segmentUrlRegex=["/ads/"]} + * 的既有规则在自检里只匹配到广告片，真实运行却会命中所有 + * {@code ?ref=/ads/} 的正片 —— 超删比例低于 35% 时 {@code fallback=false}， + * 错误不被回退兜住。 + */ +public class CompiledRuleGateTest { + + @Test + public void rejectsRuleWithUnanchoredSegmentRegex() { + String detail = "{\"id\":\"x\",\"playlistHostSuffixes\":[\"v.example.com\"]," + + "\"segmentUrlRegex\":[\"/ads/\"],\"minimumSignals\":1}"; + + assertTrue("未锚定的切片正则必须被识别", + AdFeedbackDataSource.hasUnanchoredSegmentRegex(detail)); + assertNull("不得为这类规则背书", AdFeedbackDataSource.compiledOf(detail)); + } + + @Test + public void acceptsRuleWithAnchoredSegmentRegex() { + String detail = "{\"id\":\"x\",\"playlistHostSuffixes\":[\"v.example.com\"]," + + "\"segmentUrlRegex\":[\"^[^?#]*\\\\Q/ads/\\\\E\"],\"minimumSignals\":1}"; + + assertFalse(AdFeedbackDataSource.hasUnanchoredSegmentRegex(detail)); + assertNotNull("锚定过的正则语义与自检一致，可以放行", + AdFeedbackDataSource.compiledOf(detail)); + } + + @Test + public void acceptsRuleWithoutSegmentRegexAtAll() { + // 只靠 host + 跨域的规则不涉及 query/fragment 语义差异 + String detail = "{\"id\":\"x\",\"playlistHostSuffixes\":[\"v.example.com\"]," + + "\"hostSuffixes\":[\"ad.other.com\"],\"requireCrossDomain\":true," + + "\"minimumSignals\":2}"; + + assertFalse(AdFeedbackDataSource.hasUnanchoredSegmentRegex(detail)); + assertNotNull(AdFeedbackDataSource.compiledOf(detail)); + } + + @Test + public void treatsUnparsableDetailAsUnanchored() { + assertTrue("解析不出来时保守判为未锚定", + AdFeedbackDataSource.hasUnanchoredSegmentRegex("{\"segmentUrlRegex\":[")); + assertNull(AdFeedbackDataSource.compiledOf("not json")); + assertNull(AdFeedbackDataSource.compiledOf(null)); + assertNull(AdFeedbackDataSource.compiledOf(" ")); + } + + /** compiledOf 用本地 Gson，不依赖 Application —— 纯 JVM 下必须可用。 */ + @Test + public void roundTripsWithoutApplicationContext() { + String detail = "{\"id\":\"baofeng\",\"name\":\"暴风\",\"version\":1," + + "\"playlistHostSuffixes\":[\"v.example.com\"]," + + "\"hostSuffixes\":[\"ad.other.com\"]," + + "\"minDuration\":6.3,\"maxDuration\":6.5," + + "\"requireCrossDomain\":true,\"minimumSignals\":2}"; + + HlsManifestCleaner.Rule rule = AdFeedbackDataSource.compiledOf(detail); + + assertNotNull("本地 Gson 必须能在纯 JVM 下还原规则", rule); + // 还原出的规则行为与直接构造的一致 + String manifest = manifest(); + HlsManifestCleaner.Result viaJson = HlsManifestCleaner.clean(BASE, manifest, List.of(rule)); + HlsManifestCleaner.Result direct = HlsManifestCleaner.clean(BASE, manifest, + List.of(HlsAdRule.createUserRule("baofeng", "暴风", + List.of("v.example.com"), List.of("ad.other.com"), List.of(), + 6.3, 6.5, false, true, 2).compile())); + + assertEquals(direct.removedSegments(), viaJson.removedSegments()); + assertEquals(direct.fallback(), viaJson.fallback()); + assertEquals(direct.manifest(), viaJson.manifest()); + } + + private static final String BASE = "https://v.example.com/index.m3u8"; + + private static String manifest() { + StringBuilder text = new StringBuilder("#EXTM3U\n#EXT-X-TARGETDURATION:10\n"); + for (int i = 0; i < 30; i++) { + boolean ad = i >= 10 && i < 13; + if (i == 10 || i == 13) text.append("#EXT-X-DISCONTINUITY\n"); + text.append("#EXTINF:").append(ad ? "6.400" : "8.000").append(",\n"); + text.append(ad ? "https://ad.other.com/ads/" + i + ".ts" + : "https://v.example.com/seg/" + i + ".ts").append('\n'); + } + return text.append("#EXT-X-ENDLIST\n").toString(); + } + + /** + * 端到端：带未锚定正则的既有规则不得被建议启用。 + * + *

这条规则在自检里看起来只删广告，真实运行会连 {@code ?ref=/ads/} 的正片 + * 一起删，且超删比例低于 35% 故不触发回退。 + */ + @Test + public void doesNotSuggestEnablingRuleThatOverDeletesViaQuery() { + String detail = "{\"id\":\"loose\",\"playlistHostSuffixes\":[\"v.example.com\"]," + + "\"hostSuffixes\":[\"ad.other.com\"]," + + "\"segmentUrlRegex\":[\"/ads/\"],\"minimumSignals\":1}"; + ExistingRuleClassifier.RuleState loose = new ExistingRuleClassifier.RuleState( + "vod:src:loose", "loose", "宽松规则", false, true, + List.of("ad.other.com"), AdFeedbackDataSource.compiledOf(detail)); + + List inside = new ArrayList<>(); + for (int i = 10; i < 13; i++) { + inside.add(new SegmentFact(i, "ad.other.com", "/ads/" + i + ".ts", 6.4, i == 10)); + } + List outside = new ArrayList<>(); + for (int i = 0; i < 10; i++) { + outside.add(new SegmentFact(i, "v.example.com", "/seg/" + i + ".ts", 8.0, false)); + } + for (int i = 13; i < 40; i++) { + outside.add(new SegmentFact(i, "v.example.com", "/seg/" + i + ".ts", 8.0, i == 13)); + } + AdIntervalEvidence evidence = new AdIntervalEvidence( + 80_000, 99_200, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + "v.example.com", "/play/index.m3u8", true, + inside, outside, true, true, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + + AdAttribution attribution = ExistingRuleClassifier.classify(evidence, + new ExistingRuleClassifier.Input(List.of(loose), false, List.of())); + + assertTrue("带未锚定正则的规则不得被建议启用", + attribution == null + || attribution.remediation() != RemediationKind.ENABLE_EXISTING_RULE); + } +} From 1a0f44f0917a66f31c8b62c4e7974b2e98f5aeff Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Sun, 30 Aug 2026 21:05:57 +0800 Subject: [PATCH 27/28] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E6=96=B0?= =?UTF-8?q?=E8=A7=84=E5=88=99=E5=8F=A0=E5=8A=A0=E5=B7=B2=E5=90=AF=E7=94=A8?= =?UTF-8?q?=E8=A7=84=E5=88=99=E5=90=8E=E8=B6=8A=E8=BF=87=E5=9B=9E=E9=80=80?= =?UTF-8?q?=E9=97=B8=E9=97=A8=E4=BC=9A=E8=BF=9E=E5=B8=A6=E4=B8=A2=E6=8E=89?= =?UTF-8?q?=E5=8E=9F=E6=9C=89=E5=87=80=E5=8C=96?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit HlsManifestCleaner 的三道回退闸门（全删 / 删除比例 > 35% / 删除时长 > 90s） 作用于所有规则合并后的删除总量，而自检此前只验证新规则单独生效的效果。一条 单独安全的规则叠加到已生效规则上可能越过闸门，导致整份 manifest 回退 —— 不只 新规则无效，连该站原有的结构化净化也一起丢掉，比不加规则更差。 把已启用规则接进自检：AdFeedbackDataSource.activeHlsRules 经 Host 与三个分类器 传到 RuleSelfCheck，合并后跑一遍真实 cleaner，回退即拒绝。 几个刻意的取舍： - 不追究回退是谁造成的。「已启用规则自己就回退时放行」看似合理（那时生产上已 整体不净化），实则错误：判据只对证据里这一份 manifest 成立，而规则一旦保存就 对该站所有剧集长期生效。实测同一站点另一集里已启用规则并未越界（删 10/30）， 叠加后越界回退，原本正常工作的 10 片删除全丢。且合并后回退时新规则在该 manifest 上删 0 片，对它是否有效没有任何证据。 - activeHlsRules 用 null 表示「无法确定」，与空列表（确无规则生效）区分。 removedCount 对规则集单调，于是用子集判出「会回退」是安全的，判出「不回退」 却对全集一无所知。带未锚定 segmentUrlRegex 的已启用规则在合成 manifest 上 命中 0 片（合成 URI 不含 query/fragment），此类规则存在时整体返回 null 而非 把它剔除后预测。Host.activeHlsRules 因此不给默认实现，读取失败也传播 null。 - 不做「区间外切片逐一比对」。合并后的删除集恒等于各自删除集的并集，而单独校验 已钉住「新规则只删区间内」，那种比对恒真。 顺带堵上一个既有洞：区间内外存在同一 URI 时，「删反了」（留区间内、删区间外） 与正确结果的 URI 次数和删除总数完全相同，两道判据同时被抵消而放行，真实运行删 正片且 fallback=false。真实 URL 靠 query 区分而证据已去敏，补判据无解，只能弃权。 代价是也拒掉一部分本来安全的重名场景，注释里记了权衡与将来的救回思路。 测试：新增 RuleStackingTest；补上此前完全缺失的控制器串接覆盖（Host 方法曾是 默认实现，串接断开不会有红灯）；每条新测试都用反向插桩确认非空转。 --- .../tv/ad/feedback/AdFeedbackController.java | 23 ++- .../tv/ad/feedback/AdFeedbackDataSource.java | 48 +++++ .../tv/ad/feedback/AdFeedbackHostAdapter.java | 17 ++ .../feedback/DomainReputationClassifier.java | 16 +- .../ad/feedback/ExistingRuleClassifier.java | 19 +- .../tv/ad/feedback/HlsSegmentClassifier.java | 18 +- .../android/tv/ad/feedback/RuleSelfCheck.java | 106 ++++++++++- .../ad/feedback/AdFeedbackControllerTest.java | 58 ++++++ .../feedback/AdFeedbackHostAdapterTest.java | 10 +- .../ad/feedback/RuleSelfCheckAliasTest.java | 39 +++- .../tv/ad/feedback/RuleStackingTest.java | 170 ++++++++++++++++++ 11 files changed, 504 insertions(+), 20 deletions(-) create mode 100644 app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleStackingTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java index add0e942ed..671c5904a1 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackController.java @@ -1,6 +1,7 @@ package com.fongmi.android.tv.ad.feedback; import com.fongmi.android.tv.bean.M3u8Evidence; +import com.fongmi.android.tv.utils.HlsManifestCleaner; import java.util.ArrayList; import java.util.List; @@ -46,6 +47,17 @@ public final class AdFeedbackController { /** 已知 HLS 规则状态。 */ List hlsRuleStates(); + /** + * 当前已启用的 HLS 规则，新规则要与它们叠加后仍然安全 —— + * {@code HlsManifestCleaner} 的三道回退闸门作用于合并后的删除总量。 + * + *

{@code null} 表示无法确定（存在模拟不了的已启用规则，或读取配置失败）， + * 与空列表语义不同：前者让自检拒绝新规则，后者表示站内确无规则生效。 + * 刻意不给默认实现 —— 默认返回空列表等于代替实现者宣称「没有规则生效」， + * 那是它无从知道的事，而这个方向恰好是不安全的那一侧。 + */ + List activeHlsRules(); + /** 正片保护正则。 */ List protectingExcludes(); @@ -174,14 +186,19 @@ public final class AdFeedbackController { host.context(), playlist, session.startMs(), session.endMs(), session.startOrigin(), blacklist, host.legacyHeuristicActive()); + // 已启用规则参与自检：新规则叠加上去后不得越过 cleaner 的回退闸门 + List activeRules = host.activeHlsRules(); + List attributions = new ArrayList<>(); - attributions.add(HlsSegmentClassifier.classify(evidence)); + attributions.add(HlsSegmentClassifier.classify(evidence, activeRules)); attributions.add(DomainReputationClassifier.classify(evidence, new DomainReputationClassifier.Input(blacklist, host.siteBaselineHosts(), - host.interfaceCandidateHosts(), host.interfaceSourceName()))); + host.interfaceCandidateHosts(), host.interfaceSourceName()), + activeRules)); attributions.add(ExistingRuleClassifier.classify(evidence, new ExistingRuleClassifier.Input(host.hlsRuleStates(), - host.legacyHeuristicActive(), host.protectingExcludes()))); + host.legacyHeuristicActive(), host.protectingExcludes()), + activeRules)); for (AdAttribution attribution : attributions) { if (attribution == null) diagnostics.record(AdFeedbackDiagnostics.Code.CHANNEL_ABSTAINED); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java index f88cc37f6f..b4259fd1aa 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackDataSource.java @@ -7,6 +7,7 @@ import com.fongmi.android.tv.api.config.UserAdRuleStore; import com.fongmi.android.tv.bean.HlsAdRule; import com.fongmi.android.tv.bean.ImportedAdRuleCandidate; import com.fongmi.android.tv.bean.UserAdRule; +import com.fongmi.android.tv.setting.Setting; import com.fongmi.android.tv.utils.HlsManifestCleaner; import com.google.gson.Gson; import com.google.gson.JsonArray; @@ -114,6 +115,53 @@ public final class AdFeedbackDataSource { } } + /** + * 当前已启用的 HLS 规则，供叠加校验预测回退闸门。 + * + *

返回 {@code null} 表示**无法确定**，调用方必须因此拒绝新规则 —— 不是 + * 「没有规则」。两者的区别是安全性的关键：叠加校验靠「合并后是否越过闸门」下 + * 结论，而 {@code removedCount} 对规则集单调，于是 + * + *

+ * + *

所以凡是有一条已启用规则无法被自检忠实模拟，就整体返回 {@code null}， + * 而不是把它剔除后拿剩下的规则去预测。带未锚定 {@code segmentUrlRegex} 的规则 + * 正属此类：合成 manifest 的 URI 不含 query 与 fragment（见 + * {@link #hasUnanchoredSegmentRegex}），这类规则在合成上命中 0 片，真实运行却 + * 可能命中大量切片。此前把它们静默剔除，等于让「合并后 3/30 不回退」这种结论 + * 建立在比生产小得多的规则集上，恰好放过「叠加后整份 manifest 回退」—— + * 正是叠加校验要拦的那件事。 + * + *

去广告总开关关闭时返回空列表：此时 {@code HlsAdblockPipeline} 根本不会被 + * 调用，所有规则确实删 0 片，「没有规则生效」是真话而非近似。 + * + *

刻意不按播放内核过滤：结构化净化目前只在 IJK 路径与本地 {@code /m3u8} + * 代理上执行（MPV 会为时间戳完整性丢弃净化结果，EXO/系统内核走 legacy）， + * 但内核可在播放中回退切换，而规则一旦保存就长期生效。按当次内核放宽，等于 + * 让规则的安全性取决于保存那一刻的内核，换内核后即失效。 + */ + public static List activeHlsRules() { + try { + if (!Setting.isAdblock()) return List.of(); + List simulatable = new ArrayList<>(); + for (HlsRuleConfig.Entry entry : HlsRuleConfig.getEntries()) { + if (entry == null || !entry.enabled() || !entry.valid()) continue; + HlsManifestCleaner.Rule compiled = compiledOf(entry.detail()); + // compiledOf 已内含未锚定正则的拒绝。拿不到编译产物就无法预测它 + // 的删除量，只能整体弃权 —— 剔除它再预测是不安全的方向。 + if (compiled == null) return null; + simulatable.add(compiled); + } + return List.copyOf(simulatable); + } catch (RuntimeException e) { + // 读配置失败同样是「无法确定」，不能退化成「没有规则」 + return null; + } + } + /** * 从规则详情 JSON 还原已编译的规则，供 {@code RuleSelfCheck} 实跑验证。 * diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java index 90e2c6d0cf..0beb47998f 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapter.java @@ -1,6 +1,7 @@ package com.fongmi.android.tv.ad.feedback; import com.fongmi.android.tv.bean.M3u8Evidence; +import com.fongmi.android.tv.utils.HlsManifestCleaner; import com.fongmi.android.tv.utils.M3u8Parser; import java.util.List; @@ -67,6 +68,7 @@ public final class AdFeedbackHostAdapter implements AdFeedbackController.Host { Supplier interfaceSourceName, Supplier> protectingExcludes, Supplier> hlsRuleStates, + Supplier> activeHlsRules, SitePlaylistHostBaseline baseline) { public static Deps defaults() { @@ -76,6 +78,7 @@ public final class AdFeedbackHostAdapter implements AdFeedbackController.Host { AdFeedbackDataSource::interfaceSourceName, AdFeedbackDataSource::protectingExcludes, AdFeedbackDataSource::hlsRuleStates, + AdFeedbackDataSource::activeHlsRules, SiteHostBaselineStore.get()); } } @@ -218,6 +221,20 @@ public final class AdFeedbackHostAdapter implements AdFeedbackController.Host { } } + /** + * 与其他 {@code orEmpty} 式读取不同，这里**不把失败折成空列表**：空列表意味着 + * 「站内确无规则生效」，会让自检跳过叠加校验；读不到时的真实状态是「不知道」， + * 必须原样传 {@code null} 让自检拒绝。折成空列表是 fail-open。 + */ + @Override + public List activeHlsRules() { + try { + return deps.activeHlsRules().get(); + } catch (RuntimeException e) { + return null; + } + } + @Override public List protectingExcludes() { return orEmpty(deps.protectingExcludes()); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java index 574f5928ba..29addb209d 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/DomainReputationClassifier.java @@ -1,5 +1,7 @@ package com.fongmi.android.tv.ad.feedback; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + import java.util.ArrayList; import java.util.LinkedHashSet; import java.util.List; @@ -49,8 +51,18 @@ public final class DomainReputationClassifier { } } - /** 无结论时返回 null 表示弃权。 */ + /** 无结论时返回 null 表示弃权。不含叠加校验，仅供单测。 */ public static AdAttribution classify(AdIntervalEvidence evidence, Input input) { + return classify(evidence, input, List.of()); + } + + /** + * 带叠加校验的版本：新规则要与 {@code activeRules} 合并后仍然安全。 + * + * @param activeRules 当前已启用并生效的 HLS 规则 + */ + public static AdAttribution classify(AdIntervalEvidence evidence, Input input, + List activeRules) { if (evidence == null || evidence.inside().isEmpty()) return null; Input safe = input == null ? Input.empty() : input; @@ -112,7 +124,7 @@ public final class DomainReputationClassifier { Double.NaN, Double.NaN, false, true, 2); // 最终守门：用真实 cleaner 验证只删区间内。域名条件同样挡不住 // 「广告与正片共用第三方 CDN、靠路径区分」这类场景。 - if (!RuleSelfCheck.isSafe(evidence, payload)) return null; + if (!RuleSelfCheck.isSafe(evidence, payload, activeRules)) return null; return new AdAttribution(CHANNEL_ID, AdCategory.THIRD_PARTY_CDN_SEGMENT, confidence, RiskLevel.LOW, evidenceLines, RemediationKind.HLS_STRUCTURED_RULE, payload); diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java index d000c518c5..912f391558 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/ExistingRuleClassifier.java @@ -77,12 +77,22 @@ public final class ExistingRuleClassifier { } } - /** 无诊断结论时返回 null。 */ + /** 无诊断结论时返回 null。不含叠加校验，仅供单测。 */ public static AdAttribution classify(AdIntervalEvidence evidence, Input input) { + return classify(evidence, input, List.of()); + } + + /** + * 带叠加校验的版本：待启用规则要与 {@code activeRules} 合并后仍然安全。 + * + * @param activeRules 当前已启用并生效的 HLS 规则 + */ + public static AdAttribution classify(AdIntervalEvidence evidence, Input input, + List activeRules) { if (evidence == null || evidence.inside().isEmpty()) return null; Input safe = input == null ? Input.empty() : input; - RuleState disabled = findDisabledMatch(evidence, safe.hlsRules()); + RuleState disabled = findDisabledMatch(evidence, safe.hlsRules(), activeRules); if (disabled != null) { List lines = new ArrayList<>(); lines.add(String.format(Locale.US, "已有规则「%s」覆盖该域名但当前未启用", @@ -125,7 +135,8 @@ public final class ExistingRuleClassifier { * 「广告与正片共用同一 CDN」时会命中全部切片，回退并连带停掉 legacy 启发式。 * 这条路径此前完全没有守卫。 */ - private static RuleState findDisabledMatch(AdIntervalEvidence evidence, List rules) { + private static RuleState findDisabledMatch(AdIntervalEvidence evidence, List rules, + List activeRules) { for (RuleState rule : rules) { if (rule.enabled() || !rule.valid() || rule.hostSuffixes().isEmpty()) continue; boolean covers = evidence.inside().stream() @@ -133,7 +144,7 @@ public final class ExistingRuleClassifier { if (!covers) continue; // 拿不到编译产物时无法验证，宁可不建议 if (rule.compiled() == null) continue; - if (!RuleSelfCheck.isSafe(evidence, rule.compiled())) continue; + if (!RuleSelfCheck.isSafe(evidence, rule.compiled(), activeRules)) continue; return rule; } return null; diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java index d3a561d536..29bd2096c2 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/HlsSegmentClassifier.java @@ -1,5 +1,7 @@ package com.fongmi.android.tv.ad.feedback; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + import java.util.ArrayList; import java.util.HashMap; import java.util.List; @@ -74,9 +76,21 @@ public final class HlsSegmentClassifier { } /** - * 对区间做结构归因。无法判断时返回 null 表示弃权。 + * 对区间做结构归因，不含叠加校验。无法判断时返回 null 表示弃权。 + * + *

仅供单测；生产路径走带 {@code activeRules} 的版本。 */ public static AdAttribution classify(AdIntervalEvidence evidence) { + return classify(evidence, List.of()); + } + + /** + * 带叠加校验的版本：新规则要与 {@code activeRules} 合并后仍然安全。 + * + * @param activeRules 当前已启用并生效的 HLS 规则 + */ + public static AdAttribution classify(AdIntervalEvidence evidence, + List activeRules) { if (evidence == null || !evidence.hasSegmentEvidence()) return null; // 区间已被结构化规则整体处理，用户看到的广告来自别处，不应再加 HLS 规则 if (evidence.handledByStructuredRule()) return null; @@ -92,7 +106,7 @@ public final class HlsSegmentClassifier { // 最终守门：用真实 cleaner 跑一遍，确认只删区间内、不碰区间外、不触发回退。 // 逐条件的对照校验挡不住条件间的组合效应（如 duration + crossDomain 在 // 区间外同时成立），也预判不到删除比例与时长闸门。 - if (!RuleSelfCheck.isSafe(evidence, payload)) return null; + if (!RuleSelfCheck.isSafe(evidence, payload, activeRules)) return null; return new AdAttribution(CHANNEL_ID, categoryOf(signals), confidence, RiskLevel.MEDIUM, describe(evidence, signals, payload), diff --git a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java index 9213512892..cf868c2f59 100644 --- a/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java +++ b/app/src/main/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheck.java @@ -46,26 +46,108 @@ final class RuleSelfCheck { private static final int MANIFEST_HEADER_LINES = 16; /** + * 只做单独校验，不含叠加。仅供单测直接验证 {@link #isSafeAlone} 的判据 —— + * 生产路径必须走三参版本，把已启用规则一起算进去。 + * * @return 载荷是否安全可落地 */ static boolean isSafe(AdIntervalEvidence evidence, RulePayload payload) { + return isSafe(evidence, payload, List.of()); + } + + /** 带叠加校验的载荷版本。 */ + static boolean isSafe(AdIntervalEvidence evidence, RulePayload payload, + List activeRules) { if (payload.isEmpty()) return false; HlsManifestCleaner.Rule rule = compile(payload); if (rule == null) return false; - return isSafe(evidence, rule); + return isSafe(evidence, rule, activeRules); } /** - * 直接校验一条已编译的规则，供「启用已有规则」路径复用。 + * 校验规则单独生效、以及与站内已启用规则叠加后是否都安全。 * - *

那条路径此前完全绕过自检：它只判断规则的 hostSuffixes 与区间内某片同域， - * 不看该规则的其余条件。实测一条 {@code minimumSignals=1} 的既有规则在 - * 「广告与正片共用同一 CDN」时会命中全部切片。 + *

也是「启用已有规则」路径的守门人：那条路径此前只判断规则的 hostSuffixes + * 与区间内某片同域，不看其余条件。实测一条 {@code minimumSignals=1} 的既有规则 + * 在「广告与正片共用同一 CDN」时会命中全部切片。 + * + *

{@code HlsManifestCleaner.matches} 是「任一规则命中即删」，而三道回退闸门 + * （全删 / 删除比例 > 35% / 删除时长 > 90s）作用于**合并后**的删除总量。 + * 一条单独安全的规则叠加到已生效规则上可能越过闸门，导致整份 manifest 回退 —— + * 不只新规则无效，连原有的结构化净化也一起丢掉。 + * + *

刻意不做「区间外切片逐一比对」：合并后的删除集恒等于两条规则各自删除集的 + * 并集（{@code matches} 逐片独立求值），而 {@link #isSafeAlone} 已经钉住「新规则 + * 单独只删区间内」。合并后多删的区间外切片全部来自已启用规则 —— 那是它们本来 + * 就会删的，与新规则无关。于是新规则对区间外的影响恒为零，只有闸门会因叠加而 + * 翻转，所以只查闸门。 + * + *

合并后回退即拒绝，**不追究是谁造成的**。曾经尝试过「已启用规则自己就回退 + * 时放行」，理由是那时生产上已整体不净化、没有原有净化需要保护 —— 那是错的： + * 判据只对证据里这一份 manifest 成立，而规则一旦保存就对该站所有剧集长期生效。 + * 实测同一站点另一集里已启用规则并未越界（删 10/30），叠加新规则后越界回退， + * 原本正常工作的 10 片删除全丢。更根本的是，合并后回退时新规则在这份 manifest + * 上删 0 片，我们对它是否有效**没有任何证据**，谈不上背书。 + * + *

代价是：站内只要有一条过宽的内置或接口规则处于启用，该站点的反馈功能就 + * 一直拿不到方案。这是真实的体验问题，但解法是把肇事规则指出来让用户禁用， + * 不是保存一条未经验证的规则。 + * + * @param activeRules 当前已启用且能被忠实模拟的规则。空表示站内无规则生效； + * {@code null} 表示无法确定（存在模拟不了的已启用规则，或读取 + * 配置失败），此时一律拒绝 —— 预测不了闸门就不能为规则背书。 */ - static boolean isSafe(AdIntervalEvidence evidence, HlsManifestCleaner.Rule rule) { + static boolean isSafe(AdIntervalEvidence evidence, HlsManifestCleaner.Rule rule, + List activeRules) { + if (!isSafeAlone(evidence, rule)) return false; + if (activeRules == null) return false; + if (activeRules.isEmpty()) return true; + + List ordered = ordered(evidence); + List merged = new java.util.ArrayList<>(activeRules); + merged.add(rule); + return !fallsBack(evidence, synthesize(evidence, ordered), merged); + } + + /** + * 给定规则集在合成 manifest 上是否触发回退。判不出来时按「回退」处理， + * 让调用方走拒绝路径。 + * + *

这是对**真实运行**的预测，而合成 manifest 按去敏要求丢掉了 query 与 + * fragment：靠 query 区分广告的已启用规则在这里命中不到切片，闸门预测会偏乐观。 + * {@code AdFeedbackDataSource.activeHlsRules} 因此在遇到这类规则时整体返回 + * {@code null} 而不是把它悄悄剔除 —— 剔除会让「合并后不回退」这个结论建立在 + * 比生产更小的规则集上，而那个方向不是安全的：子集不回退推不出全集不回退。 + */ + private static boolean fallsBack(AdIntervalEvidence evidence, String manifest, + List rules) { + String base = SYNTHETIC_BASE_SCHEME + evidence.playlistHost() + "/index.m3u8"; + try { + HlsManifestCleaner.Result result = HlsManifestCleaner.clean(base, manifest, rules); + return result == null || result.fallback(); + } catch (RuntimeException e) { + return true; + } + } + + /** 规则单独生效时的校验。 */ + private static boolean isSafeAlone(AdIntervalEvidence evidence, HlsManifestCleaner.Rule rule) { if (rule == null) return false; if (evidence.playlistHost().isEmpty()) return false; if (evidence.inside().isEmpty() || evidence.outside().isEmpty()) return false; + // 区间内外存在同一个 URI 时，读回判据无法区分对错，只能弃权： + // 「删了区间内的 U、留了区间外的 U」（正确）与「留了区间内的 U、删了区间外的 + // U」（删反）两种结果的 URI 次数完全相同，removedSegments 也都等于 inside 的 + // 片数 —— 两道判据同时被抵消，放行后真实运行删正片且 fallback=false， + // 用户直接看到跳帧。真实 URL 靠 query 区分（如 /hls/seg.ts?i=10 与 ?i=25）， + // 而证据按去敏要求丢掉了 query，合成 manifest 里两者塌成一个。 + // + // 代价是也拒掉一部分本来安全的重名场景（如内外同 URI 但时长不同、规则确实 + // 只命中区间内那片）。接受这个代价：判据无法自证对错时弃权是唯一安全方向， + // 而真实 playlist 的切片 URI 通常按位置唯一，内外重名基本只来自 query 区分 + // —— 误拒几乎全部落在危险场景上。若将来要救回这部分，得让合成 URI 带上 + // 位置标记（放在 query 里，锚定正则的 ^[^?#]* 跨不过 ?），不是补判据能解决的。 + if (hasUriOverlap(evidence)) return false; // 真实 manifest 的行密度高于合成，按上界预判：若真实会因规模回退， // 这条规则在播放时不会生效，不该保存。除切片行外还要留出 header @@ -102,6 +184,18 @@ final class RuleSelfCheck { return result.removedSegments() == evidence.inside().size(); } + /** 区间内外是否有切片在合成 manifest 里塌成同一个 URI。 */ + private static boolean hasUriOverlap(AdIntervalEvidence evidence) { + java.util.Set inside = new java.util.HashSet<>(); + for (SegmentFact fact : evidence.inside()) { + inside.add(uriOf(fact, evidence.playlistHost())); + } + for (SegmentFact fact : evidence.outside()) { + if (inside.contains(uriOf(fact, evidence.playlistHost()))) return true; + } + return false; + } + /** 统计净化后 manifest 里每个切片 URI 的出现次数。 */ private static java.util.Map segmentUriCounts(String cleaned) { java.util.Map counts = new java.util.HashMap<>(); diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java index bd44351bda..d72c2a941d 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackControllerTest.java @@ -6,7 +6,9 @@ import static org.junit.Assert.assertNotNull; import static org.junit.Assert.assertNull; import static org.junit.Assert.assertTrue; +import com.fongmi.android.tv.bean.HlsAdRule; import com.fongmi.android.tv.bean.M3u8Evidence; +import com.fongmi.android.tv.utils.HlsManifestCleaner; import org.junit.Test; @@ -34,6 +36,56 @@ public class AdFeedbackControllerTest { return M3u8Evidence.create(segments, List.of(3, 6), durations, switches); } + /** + * 控制器必须把 {@code host.activeHlsRules()} 传给分类器做叠加校验。 + * + *

没有这条测试，把 {@code classify} 里的 activeRules 换成 {@code List.of()} + * 整个套件依然全绿 —— 串接一旦断开不会有任何红灯。 + */ + @Test + public void passesActiveRulesIntoClassifiersForStackingCheck() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + // 先确认这份证据在无已启用规则时能产出方案 + AdFeedbackController baseline = new AdFeedbackController(host); + baseline.onMarkedInterval(30_000, 49_200); + assertTrue("前置条件：无叠加时应有可落地方案", + host.shownSessions.get(1).hasActionablePlan()); + + // 一条已启用规则删掉另外 9 片正片（下标 6-14，各 10s）。与新规则的 3 片 + // 合并后同时越过两道闸门：12/30 = 40% > 35%，且 90 + 19.2 = 109.2s > 90s。 + // 刻意让两道都越界，避免测试卡在单一闸门的边界上。 + FakeHost stacked = new FakeHost(); + stacked.playlist = adBlockPlaylist(); + stacked.activeRules = List.of(HlsAdRule.createUserRule( + "active", "active", List.of("v.example.com"), List.of(), + List.of("^[^?#]*/seg/(?:6|7|8|9|10|11|12|13|14)\\.ts"), + null, null, false, false, 1).compile()); + AdFeedbackController controller = new AdFeedbackController(stacked); + + controller.onMarkedInterval(30_000, 49_200); + + assertFalse("叠加后越过回退闸门，不该再给出可落地方案", + stacked.shownSessions.get(1).hasActionablePlan()); + } + + /** + * {@code activeHlsRules()} 返回 {@code null}（无法确定站内规则）时必须弃权， + * 而不是当成「没有规则」继续放行。 + */ + @Test + public void abstainsWhenHostCannotDetermineActiveRules() { + FakeHost host = new FakeHost(); + host.playlist = adBlockPlaylist(); + host.activeRules = null; + AdFeedbackController controller = new AdFeedbackController(host); + + controller.onMarkedInterval(30_000, 49_200); + + assertFalse("规则集未知时不该给出可落地方案", + host.shownSessions.get(1).hasActionablePlan()); + } + @Test public void showsPendingSessionBeforeAnalysisThenVerdict() { FakeHost host = new FakeHost(); @@ -296,6 +348,7 @@ public class AdFeedbackControllerTest { private boolean deferBackground; private List blacklist = List.of(); private List ruleStates = List.of(); + private List activeRules = List.of(); @Override public long positionMs() { @@ -350,6 +403,11 @@ public class AdFeedbackControllerTest { return ruleStates; } + @Override + public List activeHlsRules() { + return activeRules; + } + @Override public List protectingExcludes() { return List.of(); diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapterTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapterTest.java index 2d7e838a03..8c7dd7dc21 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapterTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/AdFeedbackHostAdapterTest.java @@ -64,6 +64,7 @@ public class AdFeedbackHostAdapterTest { () -> { throw new IllegalStateException("prefs unavailable"); }, () -> { throw new IllegalStateException("prefs unavailable"); }, () -> { throw new IllegalStateException("prefs unavailable"); }, + () -> { throw new IllegalStateException("prefs unavailable"); }, new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage())); AdFeedbackHostAdapter adapter = adapter(new FakePlayback(), throwing); @@ -73,18 +74,23 @@ public class AdFeedbackHostAdapterTest { assertEquals("", adapter.interfaceSourceName()); assertTrue(adapter.protectingExcludes().isEmpty()); assertTrue(adapter.hlsRuleStates().isEmpty()); + // 唯一的例外：已启用规则读不到时必须是 null 而非空列表 —— 空列表意味着 + // 「站内确无规则生效」，会让自检跳过叠加校验，是 fail-open。 + assertNull("读取失败必须传播为『无法确定』", adapter.activeHlsRules()); } @Test public void nullConfigValuesBecomeEmptyLists() { AdFeedbackHostAdapter.Deps nulls = new AdFeedbackHostAdapter.Deps( - () -> null, () -> null, () -> null, () -> null, () -> null, + () -> null, () -> null, () -> null, () -> null, () -> null, () -> null, new SitePlaylistHostBaseline(new SitePlaylistHostBaseline.MemoryStorage())); AdFeedbackHostAdapter adapter = adapter(new FakePlayback(), nulls); assertTrue(adapter.blacklistedHosts().isEmpty()); assertEquals("", adapter.interfaceSourceName()); assertTrue(adapter.hlsRuleStates().isEmpty()); + // activeHlsRules 的 null 是有语义的，不能被折成空列表 + assertNull("数据源的『无法确定』必须原样透传", adapter.activeHlsRules()); } @Test @@ -146,7 +152,7 @@ public class AdFeedbackHostAdapterTest { private static AdFeedbackHostAdapter.Deps deps(SitePlaylistHostBaseline baseline) { return new AdFeedbackHostAdapter.Deps( - List::of, List::of, () -> "", List::of, List::of, baseline); + List::of, List::of, () -> "", List::of, List::of, List::of, baseline); } private static final class FakePlayback implements AdFeedbackHostAdapter.Playback { diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckAliasTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckAliasTest.java index 0c72dfdbb1..70fe7b7e8a 100644 --- a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckAliasTest.java +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleSelfCheckAliasTest.java @@ -57,6 +57,36 @@ public class RuleSelfCheckAliasTest { RuleSelfCheck.isSafe(evidence(inside, outside), payload)); } + /** + * 同一 URI 同时出现在区间内与区间外时，两道判据会互相抵消。 + * + *

规则漏删区间内那片、却删掉区间外那片：URI 次数比对因两者同名而仍然相等， + * {@code removedSegments() == inside.size()} 也仍然成立，于是放行 —— 真实运行 + * 删正片且 {@code fallback=false}，错误不被回退兜住。真实 URL 靠 query 区分 + * （{@code /hls/seg.ts?i=10} 与 {@code ?i=25}），而证据按去敏要求丢掉 query， + * 合成 manifest 无法区分，只能弃权。 + */ + @Test + public void abstainsWhenTheSameUriAppearsInsideAndOutside() { + // 区间内 1 片广告，URI 为 X，时长 8.0（只满足 host 一个信号，会被漏删） + List inside = List.of( + new SegmentFact(12, "ad.other.com", "/ads/x.ts", 8.0, true)); + + List outside = new ArrayList<>(); + // 区间外同一个 URI X，时长 6.4（host + duration 两个信号，会被误删） + outside.add(new SegmentFact(5, "ad.other.com", "/ads/x.ts", 6.4, false)); + for (int i = 0; i < 10; i++) { + outside.add(new SegmentFact(20 + i, PLAYLIST_HOST, "/seg/" + i + ".ts", 8.0, false)); + } + + RulePayload payload = RulePayload.ofHlsRule( + List.of(PLAYLIST_HOST), List.of("ad.other.com"), List.of(), + 6.3, 6.5, false, false, 2); + + assertFalse("删反了（留区间内、删区间外）却因 URI 重名使判据抵消，必须拒绝", + RuleSelfCheck.isSafe(evidence(inside, outside), payload)); + } + /** 代理式内嵌 URL 的 path 会原样保留，同样能遮蔽真实删除。 */ @Test public void rejectsWhenProxyStylePathShadowsRemovedSegment() { @@ -79,7 +109,14 @@ public class RuleSelfCheckAliasTest { RuleSelfCheck.isSafe(evidence(inside, outside), payload)); } - /** 同一 URI 在 playlist 中复用（片头片尾同一 promo）不得造成误判。 */ + /** + * 同一 URI 在 playlist 中复用（片头片尾同一 promo）不得造成误判。 + * + *

现在由 {@code hasUriOverlap} 前置拦下，不再走到 URI 次数比对。保留这条 + * 测试是因为它钉的是**结果**（必须拒绝），而拒绝的理由换成更早的判据仍然正确； + * 次数比对本身由 {@link #rejectsWhenRemovedOutsideUriIsSubstringOfKeptOne} + * 等区间外不重名的用例继续覆盖。 + */ @Test public void rejectsWhenDuplicateUriAppearsInsideAndOutside() { List inside = List.of( diff --git a/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleStackingTest.java b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleStackingTest.java new file mode 100644 index 0000000000..cdbcca2f3a --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/ad/feedback/RuleStackingTest.java @@ -0,0 +1,170 @@ +package com.fongmi.android.tv.ad.feedback; + +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertNull; +import static org.junit.Assert.assertTrue; + +import com.fongmi.android.tv.bean.HlsAdRule; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + +import org.junit.Test; + +import java.util.ArrayList; +import java.util.List; + +/** + * 叠加校验：一条单独安全的规则，与站内已启用规则合并后不得越过 cleaner 的 + * 回退闸门。 + * + *

{@code HlsManifestCleaner.matches} 是「任一规则命中即删」，而三道闸门 + * （全删 / 删除比例 > 35% / 删除时长 > 90s）作用于合并后的删除总量。越过闸门 + * 会让整份 manifest 回退 —— 不只新规则无效，连原有的结构化净化也一起丢掉。 + */ +public class RuleStackingTest { + + private static final String PLAYLIST_HOST = "v.example.com"; + + private static List seg(int from, int count, String host, String path, + double duration, boolean discFirst) { + List facts = new ArrayList<>(); + for (int i = 0; i < count; i++) { + facts.add(new SegmentFact(from + i, host, path + (from + i) + ".ts", + duration, discFirst && i == 0)); + } + return facts; + } + + private static AdIntervalEvidence evidence(List inside, List outside) { + return new AdIntervalEvidence( + 80_000, 99_200, StartOrigin.USER_MARKED, + "site", "站点", "剧名", "线路", "第 1 集", + PLAYLIST_HOST, "/play/index.m3u8", true, + inside, outside, true, true, List.of(), false, List.of(), + AudioIntervalFact.unavailable(), SpeechIntervalFact.unavailable()); + } + + private static HlsManifestCleaner.Rule rule(String hostSuffix, int minimumSignals) { + return HlsAdRule.createUserRule("id-" + hostSuffix, "n", + List.of(PLAYLIST_HOST), List.of(hostSuffix), List.of(), + null, null, false, true, minimumSignals).compile(); + } + + /** + * 30 片：0-9 正片、10-12 广告 A、13-21 由已启用规则负责的广告 B、22-29 正片。 + * + *

广告 B 的时长刻意不同于广告 A：分类器生成的规则会带上 + * {@code duration 6.3–6.5} 条件，若两者时长相同，「时长 + 跨域」就凑够 + * minimumSignals=2，新规则**自己**就删掉了区间外的广告 B —— 单独校验先失败， + * 测不到叠加逻辑。这个混淆由 {@link #abstainsWhenDurationWindowAlsoMatchesOutside} + * 单独覆盖。 + */ + private static AdIntervalEvidence stackedEvidence() { + List inside = seg(10, 3, "ad-a.other.com", "/ads/", 6.4, true); + List outside = new ArrayList<>( + seg(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + // 已启用规则会删掉这 9 片 + outside.addAll(seg(13, 9, "ad-b.other.com", "/adb/", 4.0, true)); + outside.addAll(seg(22, 8, PLAYLIST_HOST, "/seg/", 8.0, true)); + return evidence(inside, outside); + } + + @Test + public void rejectsNewRuleWhenCombinedRemovalCrossesRatioGate() { + // 单独删 3/30 = 10%，与已启用规则的 9 片合并后 12/30 = 40% > 35% + AdIntervalEvidence evidence = stackedEvidence(); + HlsManifestCleaner.Rule active = rule("ad-b.other.com", 2); + + assertTrue("新规则单独是安全的", + RuleSelfCheck.isSafe(evidence, rule("ad-a.other.com", 2), List.of())); + assertFalse("与已启用规则叠加后越过 35% 闸门，必须拒绝", + RuleSelfCheck.isSafe(evidence, rule("ad-a.other.com", 2), List.of(active))); + } + + @Test + public void acceptsNewRuleWhenCombinedRemovalStaysUnderGate() { + // 已启用规则只负责 2 片，合并后 5/30 = 16.7% + List inside = seg(10, 3, "ad-a.other.com", "/ads/", 6.4, true); + List outside = new ArrayList<>( + seg(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + outside.addAll(seg(13, 2, "ad-b.other.com", "/adb/", 6.4, true)); + outside.addAll(seg(15, 15, PLAYLIST_HOST, "/seg/", 8.0, true)); + AdIntervalEvidence evidence = evidence(inside, outside); + + assertTrue("合并后仍在闸门内，应放行", + RuleSelfCheck.isSafe(evidence, rule("ad-a.other.com", 2), + List.of(rule("ad-b.other.com", 2)))); + } + + /** + * 已启用规则**自己**就让 manifest 回退时，依然拒绝。 + * + *

曾经在这里放行，理由是「那时生产上已整体不净化、没有原有净化需要保护」。 + * 那是错的：判据只对证据里这一份 manifest 成立，而规则一旦保存就对该站所有 + * 剧集长期生效 —— 同一站点另一集里已启用规则可能并未越界，叠加新规则后越界 + * 回退，原本正常工作的删除全丢。更根本的是，合并后回退时新规则在这份 manifest + * 上删 0 片，对它是否有效没有任何证据，谈不上背书。 + */ + @Test + public void rejectsEvenWhenActiveRulesThemselvesAlreadyCauseFallback() { + AdIntervalEvidence evidence = stackedEvidence(); + // 一条命中全站同域切片的已启用规则：18/30 越过 35% 闸门，单独就会回退 + HlsManifestCleaner.Rule overbroad = HlsAdRule.createUserRule( + "wide", "wide", List.of(PLAYLIST_HOST), List.of(PLAYLIST_HOST), + List.of(), null, null, false, false, 1).compile(); + + assertFalse("合并后回退即拒绝，不追究是谁造成的", + RuleSelfCheck.isSafe(evidence, rule("ad-a.other.com", 2), List.of(overbroad))); + } + + /** + * {@code activeRules} 为 {@code null} 表示「无法确定站内有哪些规则生效」， + * 必须拒绝 —— 与空列表（确无规则生效）语义相反。 + * + *

数据源在遇到模拟不了的已启用规则、或读配置失败时返回 null。若把它当成 + * 空列表处理，叠加校验就被整体跳过，是 fail-open。 + */ + @Test + public void rejectsWhenActiveRulesAreUnknown() { + assertFalse("规则集未知时不能为新规则背书", + RuleSelfCheck.isSafe(stackedEvidence(), rule("ad-a.other.com", 2), null)); + } + + /** + * 分类器生成的时长窗口会把区间外同时长的跨域切片一起删掉时，必须弃权。 + * + *

这不是叠加问题、也不钉本次改动 —— 单独一条规则就已经越界： + * {@code duration 6.3–6.5} 与 {@code requireCrossDomain} 凑够 minimumSignals=2， + * 另一段广告（本应交给别的规则或保持原样）落进同一窗口就被误删。刻意传 + * {@code List.of()} 并让删除比例停在闸门内（5/30），从而钉住判据来自 + * {@code isSafeAlone} 的区间外 URI 次数比对，而不是回退闸门。 + * + *

留在这个文件里是因为它解释了 {@link #stackedEvidence} 为何要把广告 B 的 + * 时长与广告 A 错开 —— 否则叠加用例会先被这条判据拦掉，测不到闸门。 + */ + @Test + public void abstainsWhenDurationWindowAlsoMatchesOutside() { + List inside = seg(10, 3, "ad-a.other.com", "/ads/", 6.4, true); + List outside = new ArrayList<>( + seg(0, 10, PLAYLIST_HOST, "/seg/", 8.0, false)); + // 与区间内同时长、同为跨域，但域名与路径都不同 + outside.addAll(seg(13, 2, "ad-b.other.com", "/adb/", 6.4, true)); + outside.addAll(seg(15, 15, PLAYLIST_HOST, "/seg/", 8.0, true)); + + assertNull("时长窗口会连带删除区间外切片，必须弃权", + HlsSegmentClassifier.classify(evidence(inside, outside), List.of())); + } + + /** 端到端：分类器在叠加会越界时整体弃权。 */ + @Test + public void classifierAbstainsWhenStackingCrossesGate() { + AdIntervalEvidence evidence = stackedEvidence(); + List active = List.of(rule("ad-b.other.com", 2)); + + AdAttribution alone = HlsSegmentClassifier.classify(evidence, List.of()); + AdAttribution stacked = HlsSegmentClassifier.classify(evidence, active); + + assertTrue("无叠加时应能产出方案", alone != null && alone.actionable()); + assertTrue("叠加越界时必须弃权或不可落地", + stacked == null || !stacked.actionable()); + } +} From 1d611acd5410ba4dabe07a6decbab65263a005c6 Mon Sep 17 00:00:00 2001 From: Silent1566 <245409147+Silent1566@users.noreply.github.com> Date: Mon, 31 Aug 2026 17:38:02 +0800 Subject: [PATCH 28/28] =?UTF-8?q?=E6=94=AF=E6=8C=81=EF=BC=9A=E9=BB=98?= =?UTF-8?q?=E8=AE=A4=20EXO=20=E5=86=85=E6=A0=B8=E6=89=A7=E8=A1=8C=E7=BB=93?= =?UTF-8?q?=E6=9E=84=E5=8C=96=E5=8E=BB=E5=B9=BF=E5=91=8A=E8=A7=84=E5=88=99?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 此前 HlsManifestCleaner 只在 IJK 的 MpvHlsProxy 与 nano /m3u8 代理上执行， EXO（默认内核）与系统内核完全不跑结构化规则 —— 用户和接口下发的 HLS 规则在 多数用户的设备上是死的，只有 fork 版 media3 里的 legacy 启发式在工作。 新增 HlsPlaylistCleaningDataSource 挂在 MediaSourceFactory 的数据源链上： 嗅探 #EXTM3U 识别 playlist、缓冲整份、跑 HlsAdblockPipeline、交付净化结果。 放在缓存之上，好让改规则下次加载即生效而不必等淘汰。 三件必须一起做的事，否则接上比不接更糟： - 关掉 fork parser 的启发式（ExoUtil setAdblock(false)），由 pipeline 保证 两套引擎互斥。HlsAdsParser.findMinorityGroup 的判据是「最大分组须超切片总数 50%」且删除量无上限，而结构化净化会缩小它比对的分母。实测正片分布在两个 CDN （260+240 片）+ 广告 25 片时，原始 playlist 上启发式弃权（520<=525），净化后 520>500 闸门打开，删掉那 240 片正片 —— 48% 的正片静默消失。 - 只处理 VOD（#EXT-X-ENDLIST 判据，与 MpvHlsProxy.isVodPlaylist 一致）。 cleaner 会按删掉的头部片数抬高 #EXT-X-MEDIA-SEQUENCE，但它的回退闸门会在 相邻两次 reload 间翻转，回退时发出未抬高的原始序号。滑动窗口里新广告块总是 先出现在尾部（删除点不连续 → 回退），随窗口滑动才变头部连续，于是序号来回跳。 media3 的 isNewerThan 会静默丢弃「变旧」的 reload，targetDuration×3.5 后抛 PlaylistStuckException 卡播。判据刻意大小写敏感，与 cleaner 保持一致 —— 不一致会让门判 VOD 放行而 cleaner 判直播走抬高分支，缺陷原地复活。 - 隐式 IV 的加密流两套引擎都不许动（新增 HlsEncryptionGuard，放在 HlsAdblockPipeline 入口）。#EXT-X-KEY 省略 IV= 时解密 IV 由每片的 media sequence 推导，删片会让后续每片 IV 错位，画面直接花屏。这道检查曾放在 HlsManifestCleaner 里，那是错的：cleaner 唯一的拒绝方式是 Result.fallback， 而 fallback 恰是 pipeline 判定「最该由启发式接手」的信号 —— 等于把 playlist 递给启发式，IV 错位一分不少还多删 200 片正片。判据逐字镜像 media3 的 REGEX_METHOD/REGEX_IV（大小写敏感、锚定属性名）、TAG_KEY 不含冒号、按 \r\n|\n|\r 切行，覆盖 URI 里出现 ?method=none / ?iv=1、KEYFORMAT 里出现 iv=、标签与冒号间有空格、CR-only 换行这五种会骗过朴素子串匹配的形状。 测试 27 条，每条都用反向插桩确认非空转。夹具刻意做成 legacy 启发式会弃权的 均匀命名 —— 此前用独立广告域名，启发式的输出与结构化引擎逐字节相同，导致 4 条测试把规则换成 List.of() 也照样绿。另外本模块没配 returnDefaultValues， 未 mock 的 android.util.Log 会让 HlsAdsParser 第一句就抛而被 pipeline 吞掉， 启发式实际从未执行过；用 @BeforeClass 关掉 media3 日志让它真的跑起来， 并 @AfterClass 还原（进程级静态状态，无 fork 会泄漏给别的测试类）。 已知覆盖缺口：ExoUtil 的 setAdblock(false) 与 getDataSourceFactory 的挂载点 纯 JVM 都构造不出来，改回去不会有红灯，只能靠注释与人工复查。 --- .../fongmi/android/tv/player/exo/ExoUtil.java | 10 +- .../exo/HlsPlaylistCleaningDataSource.java | 324 ++++++++++ .../tv/player/exo/MediaSourceFactory.java | 17 +- .../android/tv/utils/HlsAdblockPipeline.java | 13 + .../android/tv/utils/HlsEncryptionGuard.java | 59 ++ .../HlsPlaylistCleaningDataSourceTest.java | 587 ++++++++++++++++++ .../tv/utils/HlsEncryptionGuardTest.java | 228 +++++++ 7 files changed, 1235 insertions(+), 3 deletions(-) create mode 100644 app/src/main/java/com/fongmi/android/tv/player/exo/HlsPlaylistCleaningDataSource.java create mode 100644 app/src/main/java/com/fongmi/android/tv/utils/HlsEncryptionGuard.java create mode 100644 app/src/test/java/com/fongmi/android/tv/player/exo/HlsPlaylistCleaningDataSourceTest.java create mode 100644 app/src/test/java/com/fongmi/android/tv/utils/HlsEncryptionGuardTest.java diff --git a/app/src/main/java/com/fongmi/android/tv/player/exo/ExoUtil.java b/app/src/main/java/com/fongmi/android/tv/player/exo/ExoUtil.java index a17c39de27..9788828447 100644 --- a/app/src/main/java/com/fongmi/android/tv/player/exo/ExoUtil.java +++ b/app/src/main/java/com/fongmi/android/tv/player/exo/ExoUtil.java @@ -78,7 +78,6 @@ import com.fongmi.android.tv.setting.PlaybackPerformanceCatalog; import com.fongmi.android.tv.setting.PlaybackPerformanceSetting; import com.fongmi.android.tv.setting.PlayerSetting; import com.fongmi.android.tv.subtitle.RealtimeSubtitleBufferSizeProvider; -import com.fongmi.android.tv.setting.Setting; import com.fongmi.android.tv.utils.ResUtil; import com.fongmi.android.tv.utils.UrlUtil; import com.github.catvod.crawler.SpiderDebug; @@ -249,7 +248,14 @@ public class ExoUtil { builder.setDrmConfiguration(buildDrmConfig(spec.getDrm())); builder.setRequestMetadata(buildRequestMetadata(url, headers)); builder.setMediaMetadata(spec.getMetadata()); - builder.setAdblock(Setting.isAdblock()); + // Deliberately false: ad removal for Exo lives in HlsPlaylistCleaningDataSource, + // which runs both engines through HlsAdblockPipeline. This flag would make the + // fork's HlsPlaylistParser run HlsAdsParser again on the already-cleaned text, + // and that heuristic is not independent of what precedes it — its "largest group + // exceeds 50%" gate is measured against the segment count, so removing ads can + // open a gate that was closed and make it delete main content without bound. + // See HlsPlaylistCleaningDataSource's class javadoc for the measured case. + builder.setAdblock(false); builder.setMimeType(spec.getFormat()); builder.setImageDurationMs(15000); builder.setMediaId(spec.getKey()); diff --git a/app/src/main/java/com/fongmi/android/tv/player/exo/HlsPlaylistCleaningDataSource.java b/app/src/main/java/com/fongmi/android/tv/player/exo/HlsPlaylistCleaningDataSource.java new file mode 100644 index 0000000000..74208191ed --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/player/exo/HlsPlaylistCleaningDataSource.java @@ -0,0 +1,324 @@ +package com.fongmi.android.tv.player.exo; + +import android.net.Uri; +import android.text.TextUtils; + +import androidx.annotation.NonNull; +import androidx.annotation.Nullable; +import androidx.media3.common.C; +import androidx.media3.datasource.DataSource; +import androidx.media3.datasource.DataSpec; +import androidx.media3.datasource.TransferListener; + +import com.fongmi.android.tv.api.config.HlsRuleConfig; +import com.fongmi.android.tv.setting.Setting; +import com.fongmi.android.tv.utils.HlsAdblockPipeline; +import com.fongmi.android.tv.utils.HlsManifestCleaner; +import com.github.catvod.crawler.SpiderDebug; + +import java.io.ByteArrayOutputStream; +import java.io.IOException; +import java.nio.charset.StandardCharsets; +import java.util.List; +import java.util.Map; +import java.util.function.BooleanSupplier; +import java.util.function.Supplier; + +/** + * Applies the HLS ad removal pipeline to playlist responses on the Exo path. + * + *

Exo has no local HLS proxy, so {@link HlsManifestCleaner} never ran there — + * the only ad removal was the legacy {@code HlsAdsParser} heuristic invoked inside + * the forked {@code HlsPlaylistParser}. User and interface rules were therefore + * silently inert on the default kernel. Cleaning here gives Exo the same + * {@link HlsAdblockPipeline} policy that IJK already gets through its proxy. + * + *

This class owns the whole pipeline: {@code ExoUtil} deliberately stops + * passing {@code adblock=true} into the {@code MediaItem}, so the fork's heuristic no + * longer runs behind our back. That is not tidiness — chaining them is actively + * destructive. {@code HlsAdsParser.findMinorityGroup} gates on "largest group must + * exceed 50% of segments" with no cap on how much it may delete, and removing ad + * segments shrinks the total it measures against. Measured: main content split across + * two CDN hosts (260 + 240 segments) plus 25 ads is left untouched by the heuristic on + * the raw playlist (520 <= 525), but after structured cleaning 520 > 500 opens the + * gate and it deletes the 240-segment group — 48% of the film, silently. Dropping the + * ad host can also collapse two structural groups into one, which pushes input into + * {@code findAdsByPrefixAnalysis}, a path it could never have reached. Running the two + * engines through {@link HlsAdblockPipeline} instead keeps them mutually exclusive: + * the heuristic only ever sees a manifest the structured engine did not change. + * + *

VOD only, gated on {@code #EXT-X-ENDLIST}. Live playlists must not be cleaned + * here: {@link HlsManifestCleaner} bumps {@code #EXT-X-MEDIA-SEQUENCE} by the number + * of removed head segments, but its fallback gates (removal ratio, removed duration, + * and the requirement that removals be contiguous at the head) flip between + * consecutive reloads of a sliding window — a new ad break always appears at the + * window tail first, which is non-contiguous and therefore falls back to the raw + * sequence. The emitted sequence then oscillates, media3's + * {@code HlsMediaPlaylist.isNewerThan} silently discards the "older" reload, and after + * {@code targetDuration * 3.5} playback dies with {@code PlaylistStuckException}. + * The same gate is why {@code MpvHlsProxy.applyAdblock} checks {@code isVodPlaylist}. + * + *

Sits above the cache so that editing a rule takes effect on the next load + * instead of waiting for eviction. + * + *

Playlists are recognised by sniffing the {@code #EXTM3U} header rather than by + * URL shape: one factory serves both manifest and segment loads, and media3 gives + * no reliable data-type hint at this layer. Non-playlist responses only pay for a + * {@value #SNIFF_BYTES}-byte lookahead. + */ +final class HlsPlaylistCleaningDataSource implements DataSource { + + private static final byte[] PLAYLIST_HEADER = "#EXTM3U".getBytes(StandardCharsets.US_ASCII); + /** {@code String.trim} does not strip U+FEFF, so it has to go before cleaning. */ + private static final String BOM = ""; + /** Room for a byte order mark, stray leading whitespace and the header itself. */ + private static final int SNIFF_BYTES = 64; + /** + * Mirrors {@code HlsManifestCleaner}'s own 2 MiB ceiling: past that the cleaner + * bails out anyway, so buffering the response would cost memory for nothing. + */ + private static final int MAX_PLAYLIST_BYTES = 2 * 1024 * 1024; + private static final int CHUNK_BYTES = 8 * 1024; + + private final DataSource upstream; + private final Supplier> rules; + private final BooleanSupplier adblockEnabled; + + @Nullable private byte[] buffer; + private int bufferLength; + private int bufferPos; + + HlsPlaylistCleaningDataSource(DataSource upstream) { + this(upstream, HlsPlaylistCleaningDataSource::configuredRules, Setting::isAdblock); + } + + HlsPlaylistCleaningDataSource(DataSource upstream, + Supplier> rules, + BooleanSupplier adblockEnabled) { + this.upstream = upstream; + this.rules = rules; + this.adblockEnabled = adblockEnabled; + } + + /** + * Reading the rule config must never break playback. Degrading to an empty rule + * set still lets the legacy heuristic run, which is what happened before this + * class existed. + */ + private static List configuredRules() { + try { + return HlsRuleConfig.getRules(); + } catch (RuntimeException e) { + return List.of(); + } + } + + @Override + public void addTransferListener(@NonNull TransferListener transferListener) { + upstream.addTransferListener(transferListener); + } + + @Override + public long open(@NonNull DataSpec dataSpec) throws IOException { + return open(dataSpec, dataSpec.uri == null ? "" : dataSpec.uri.toString(), + dataSpec.position, dataSpec.length); + } + + /** + * The request fields are passed in rather than read off {@code dataSpec} so that + * unit tests can drive this without constructing a {@link DataSpec} — + * {@code Uri.parse} is an unmocked android.jar stub on the JVM. + */ + long open(@Nullable DataSpec dataSpec, String url, long position, long length) + throws IOException { + buffer = null; + bufferLength = 0; + bufferPos = 0; + long upstreamLength = upstream.open(dataSpec); + if (!canInspect(url, position, length)) return upstreamLength; + + byte[] head = new byte[SNIFF_BYTES]; + int headLength = readUpTo(head); + if (!startsWithPlaylistHeader(head, headLength)) { + // Not a playlist: hand back the bytes already consumed, then stream on. + buffer = head; + bufferLength = headLength; + return upstreamLength; + } + + byte[] raw = readAll(head, headLength); + // Oversized: readAll already parked what it read in buffer; pass the rest through. + if (raw == null) return upstreamLength; + + String text = new String(raw, StandardCharsets.UTF_8); + String cleaned = cleanedOrNull(url, text); + buffer = cleaned == null ? raw : cleaned.getBytes(StandardCharsets.UTF_8); + bufferLength = buffer.length; + // The response was drained, so the served length is exactly what we hold. + // Reporting the upstream length would look like a truncated response to Exo, + // because cleaning only ever makes the playlist shorter. + return bufferLength; + } + + @Override + public int read(@NonNull byte[] target, int offset, int length) throws IOException { + if (length == 0) return 0; + if (buffer != null && bufferPos < bufferLength) { + int count = Math.min(length, bufferLength - bufferPos); + System.arraycopy(buffer, bufferPos, target, offset, count); + bufferPos += count; + return count; + } + return upstream.read(target, offset, length); + } + + @Nullable + @Override + public Uri getUri() { + return upstream.getUri(); + } + + @NonNull + @Override + public Map> getResponseHeaders() { + return upstream.getResponseHeaders(); + } + + @Override + public void close() throws IOException { + buffer = null; + bufferLength = 0; + bufferPos = 0; + upstream.close(); + } + + /** + * Only whole-response reads can be rewritten: a ranged request addresses byte + * offsets in the original body, and cleaning changes the length. + */ + private boolean canInspect(String url, long position, long length) { + if (position != 0 || length != C.LENGTH_UNSET) return false; + // The nano /m3u8 proxy already cleaned its output; doing it twice would apply + // the rules to the same playlist again, doubling the removal count and + // possibly crossing the cleaner's fallback gates. + if (HlsAdblockPipeline.isCoreM3u8Proxy(url)) return false; + try { + return adblockEnabled.getAsBoolean(); + } catch (RuntimeException e) { + return false; + } + } + + /** Fills {@code target}, stopping early only at end of input. */ + private int readUpTo(byte[] target) throws IOException { + int total = 0; + while (total < target.length) { + int read = upstream.read(target, total, target.length - total); + if (read == C.RESULT_END_OF_INPUT) break; + total += read; + } + return total; + } + + /** + * Reads the rest of the response, or returns null when it outgrows + * {@link #MAX_PLAYLIST_BYTES} — in that case the bytes consumed so far are left + * in {@link #buffer} so the caller can pass the response through untouched. + */ + @Nullable + private byte[] readAll(byte[] head, int headLength) throws IOException { + ByteArrayOutputStream out = new ByteArrayOutputStream(Math.max(headLength, 1024)); + out.write(head, 0, headLength); + byte[] chunk = new byte[CHUNK_BYTES]; + while (true) { + int read = upstream.read(chunk, 0, chunk.length); + if (read == C.RESULT_END_OF_INPUT) return out.toByteArray(); + out.write(chunk, 0, read); + if (out.size() > MAX_PLAYLIST_BYTES) { + buffer = out.toByteArray(); + bufferLength = buffer.length; + return null; + } + } + } + + /** + * @return the cleaned playlist, or null when nothing changed. + * + *

Runs the full pipeline ({@code legacyFallback=true}) because the fork's own + * heuristic pass is switched off for Exo — see the class javadoc. The pipeline + * keeps the two engines mutually exclusive: the heuristic runs only when the + * structured rules did not change the manifest. + * + *

A leading byte order mark is dropped before cleaning: + * {@code HlsManifestCleaner} requires {@code trim()} to reveal {@code #EXTM3U}, + * and {@code String.trim} does not strip U+FEFF, so a BOM would make it bail out + * on a playlist this class already recognised. Serving the cleaned text without + * the BOM is safe — the playlist parser discards it either way — and when nothing + * is cleaned the original bytes are served untouched. + */ + @Nullable + private String cleanedOrNull(String url, String text) { + String payload = text.startsWith(BOM) ? text.substring(BOM.length()) : text; + // Live windows must keep their upstream media sequence; see class javadoc. + if (!isVodPlaylist(payload)) return null; + try { + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(url, payload, rules.get(), true); + if (outcome.manifest() == null || TextUtils.equals(outcome.manifest(), payload)) return null; + SpiderDebug.log("exo-adblock", "filtered bytes=%d->%d structured=%s legacy=%s removed=%d", + text.length(), outcome.manifest().length(), outcome.structured(), + outcome.legacy(), outcome.removedSegments()); + return outcome.manifest(); + } catch (RuntimeException e) { + return null; + } + } + + /** + * Same criterion as {@code MpvHlsProxy.isVodPlaylist}: a media playlist without + * {@code #EXT-X-ENDLIST} is a live window and must not be rewritten. + * + *

Case-sensitive on purpose, matching {@code HlsManifestCleaner} and media3's + * own parser ({@code line.equals(TAG_ENDLIST)}). Accepting a lowercase spelling + * here would be worse than rejecting it: this gate would classify the playlist as + * VOD and hand it over, while the cleaner would still see it as live, take the + * media-sequence bump branch, and reintroduce exactly the cross-reload sequence + * regression the gate exists to prevent. + * + *

A multivariant playlist has no {@code #EXT-X-ENDLIST} either and is therefore + * skipped too, which is correct — it carries no segments to remove, and its + * variants are fetched as separate media playlists that do reach this code. + */ + private static boolean isVodPlaylist(String text) { + return text.contains("#EXT-X-ENDLIST"); + } + + private static boolean startsWithPlaylistHeader(byte[] data, int length) { + int index = 0; + // Skip the UTF-8 byte order mark, which the playlist parser also tolerates. + if (length >= 3 && (data[0] & 0xFF) == 0xEF && (data[1] & 0xFF) == 0xBB + && (data[2] & 0xFF) == 0xBF) index = 3; + while (index < length && Character.isWhitespace(data[index])) index++; + if (length - index < PLAYLIST_HEADER.length) return false; + for (int i = 0; i < PLAYLIST_HEADER.length; i++) { + if (data[index + i] != PLAYLIST_HEADER[i]) return false; + } + return true; + } + + static final class Factory implements DataSource.Factory { + + private final DataSource.Factory upstream; + + Factory(DataSource.Factory upstream) { + this.upstream = upstream; + } + + @NonNull + @Override + public DataSource createDataSource() { + return new HlsPlaylistCleaningDataSource(upstream.createDataSource()); + } + } +} diff --git a/app/src/main/java/com/fongmi/android/tv/player/exo/MediaSourceFactory.java b/app/src/main/java/com/fongmi/android/tv/player/exo/MediaSourceFactory.java index 20ee84c4d8..b941339893 100644 --- a/app/src/main/java/com/fongmi/android/tv/player/exo/MediaSourceFactory.java +++ b/app/src/main/java/com/fongmi/android/tv/player/exo/MediaSourceFactory.java @@ -227,12 +227,27 @@ public class MediaSourceFactory implements MediaSource.Factory { private DataSource.Factory getDataSourceFactory() { if (dataSourceFactory == null) { DataSource.Factory cacheDataSource = getCacheDataSource(new DefaultDataSource.Factory(App.get(), getHttpDataSourceFactory())); - DataSource.Factory trackedDataSource = new PlaybackBytePositionDataSource.Factory(cacheDataSource); + DataSource.Factory trackedDataSource = new PlaybackBytePositionDataSource.Factory(withPlaylistCleaning(cacheDataSource)); dataSourceFactory = new PriorityTaskDataSource.Factory(trackedDataSource, PLAYBACK_PRIORITY_MANAGER, C.PRIORITY_PLAYBACK, false); } return dataSourceFactory; } + /** + * Applies the structured HLS ad rules to playlist responses. + * + *

Exo has no local HLS proxy, so {@code HlsManifestCleaner} never ran on this + * kernel and user/interface rules were silently inert — only the legacy heuristic + * inside the forked playlist parser did anything. Wrapping here gives Exo the same + * rule engine IJK already gets through {@code MpvHlsProxy}. + * + *

Deliberately above the cache: rules must apply to cached playlists too, and + * editing a rule should take effect on the next load rather than after eviction. + */ + static DataSource.Factory withPlaylistCleaning(DataSource.Factory upstream) { + return new HlsPlaylistCleaningDataSource.Factory(upstream); + } + private CacheDataSource.Factory getCacheDataSource(DataSource.Factory upstreamFactory) { return new CacheDataSource.Factory() .setCache(getCache()) diff --git a/app/src/main/java/com/fongmi/android/tv/utils/HlsAdblockPipeline.java b/app/src/main/java/com/fongmi/android/tv/utils/HlsAdblockPipeline.java index 23590d51f3..1de773e521 100644 --- a/app/src/main/java/com/fongmi/android/tv/utils/HlsAdblockPipeline.java +++ b/app/src/main/java/com/fongmi/android/tv/utils/HlsAdblockPipeline.java @@ -2,6 +2,8 @@ package com.fongmi.android.tv.utils; import androidx.media3.exoplayer.hls.playlist.HlsAdsParser; +import com.github.catvod.crawler.SpiderDebug; + import java.net.URI; import java.util.List; @@ -10,6 +12,17 @@ public final class HlsAdblockPipeline { private HlsAdblockPipeline() {} public static Outcome apply(String url, String manifest, List rules, boolean legacyFallback) { + // Encrypted playlists whose IV is implied by the media sequence number must not be + // rewritten by *either* engine, so the check lives here rather than inside one of + // them. Putting it in HlsManifestCleaner was actively harmful: the cleaner's only + // way to decline is Result.fallback, and fallback is precisely what invites the + // legacy heuristic in below — measured, that turned "25 ads removed" into "200 main + // segments deleted by the heuristic, and every retained segment decrypting with the + // wrong IV". Declining here stops both. + if (HlsEncryptionGuard.hasImpliedSegmentIv(manifest)) { + SpiderDebug.log("hls-adblock", "declined reason=implied-segment-iv url=%s", url); + return new Outcome(manifest, false, false, 0, 0); + } HlsManifestCleaner.Result clean = HlsManifestCleaner.clean(url, manifest, rules); if (clean.changed()) { return new Outcome(clean.manifest(), true, false, clean.removedSegments(), clean.removedDurationSec()); diff --git a/app/src/main/java/com/fongmi/android/tv/utils/HlsEncryptionGuard.java b/app/src/main/java/com/fongmi/android/tv/utils/HlsEncryptionGuard.java new file mode 100644 index 0000000000..4cb2ae5979 --- /dev/null +++ b/app/src/main/java/com/fongmi/android/tv/utils/HlsEncryptionGuard.java @@ -0,0 +1,59 @@ +package com.fongmi.android.tv.utils; + +import java.util.regex.Matcher; +import java.util.regex.Pattern; + +/** + * Detects HLS playlists whose segment decryption IV is implied by the media sequence + * number, which makes them unsafe to rewrite by any means that removes segments. + * + *

When {@code #EXT-X-KEY} enables encryption but omits {@code IV=}, RFC 8216 §5.2 says + * the IV is the segment's media sequence number. Removing a segment renumbers every + * segment after it, so each one is then decrypted with the wrong IV and the stream turns + * to garbage — not a mis-cut, a black or corrupt picture. + * + *

The predicates below mirror {@code media3}'s forked {@code HlsPlaylistParser} + * exactly, because the only thing that matters is what that parser will conclude + * from the same bytes. Two of its properties are easy to get wrong, and both were: + * + *

+ */ +final class HlsEncryptionGuard { + + /** {@code media3 HlsPlaylistParser.TAG_KEY} — deliberately without the trailing colon. */ + private static final String TAG_KEY = "#EXT-X-KEY"; + /** Mirrors {@code media3 HlsPlaylistParser.REGEX_METHOD}. */ + private static final Pattern REGEX_METHOD = Pattern.compile( + "METHOD=(NONE|AES-128|SAMPLE-AES|SAMPLE-AES-CENC|SAMPLE-AES-CTR)\\s*(?:,|$)"); + /** Mirrors {@code media3 HlsPlaylistParser.REGEX_IV}. */ + private static final Pattern REGEX_IV = Pattern.compile("IV=([^,.*]+)"); + + private HlsEncryptionGuard() { + } + + /** + * @return whether any key line enables encryption without pinning an explicit IV + */ + static boolean hasImpliedSegmentIv(String manifest) { + if (manifest == null) return false; + // Split the way BufferedReader.readLine does — a lone CR ends a line too. Splitting + // on "\n" alone collapses a CR-only playlist into one giant line, hiding its keys. + for (String line : manifest.split("\r\n|\n|\r", -1)) { + if (!line.startsWith(TAG_KEY)) continue; + Matcher method = REGEX_METHOD.matcher(line); + // No recognisable METHOD: media3's parseStringAttr throws and the whole playlist + // fails to parse, so it is not something we could have corrupted. Ignore it. + if (!method.find()) continue; + if ("NONE".equals(method.group(1))) continue; + if (!REGEX_IV.matcher(line).find()) return true; + } + return false; + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/player/exo/HlsPlaylistCleaningDataSourceTest.java b/app/src/test/java/com/fongmi/android/tv/player/exo/HlsPlaylistCleaningDataSourceTest.java new file mode 100644 index 0000000000..7f8bb4fb52 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/player/exo/HlsPlaylistCleaningDataSourceTest.java @@ -0,0 +1,587 @@ +package com.fongmi.android.tv.player.exo; + +import static org.junit.Assert.assertArrayEquals; +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import android.net.Uri; + +import androidx.media3.common.C; +import androidx.media3.common.util.Log; +import androidx.media3.datasource.DataSource; +import androidx.media3.datasource.DataSpec; +import androidx.media3.datasource.TransferListener; +import androidx.media3.exoplayer.hls.playlist.HlsAdsParser; + +import com.fongmi.android.tv.bean.HlsAdRule; +import com.fongmi.android.tv.utils.HlsAdblockPipeline; +import com.fongmi.android.tv.utils.HlsManifestCleaner; + +import org.junit.AfterClass; +import org.junit.BeforeClass; +import org.junit.Test; + +import java.io.ByteArrayOutputStream; +import java.io.IOException; +import java.nio.charset.StandardCharsets; +import java.util.HashSet; +import java.util.List; +import java.util.Locale; +import java.util.Map; +import java.util.Set; + +/** + * Exo 路径上的去广告。此前 {@code HlsManifestCleaner} 只在 IJK 的本地代理里执行， + * 默认内核（EXO）上用户与接口规则完全不生效，只有 fork 版 parser 里的 legacy + * 启发式在跑。 + */ +public class HlsPlaylistCleaningDataSourceTest { + + /** + * 让 legacy 启发式在纯 JVM 下真的跑起来。 + * + *

本模块没有配 {@code testOptions { unitTests.returnDefaultValues }}，所以 + * {@code android.util.Log} 是未 mock 的桩，{@code HlsAdsParser.process} 第一句 + * {@code Log.d} 就抛 {@code RuntimeException}，而 {@code HlsAdblockPipeline} 会 + * {@code catch (Throwable)} 后原样返回 —— 启发式**从未真正执行**。上一轮的测试 + * 因此有几条钉的是这个 harness 假象而不是代码：例如「无规则时逐字节透传」在真机上 + * 是假的，启发式会删掉那 3 片广告。 + * + *

media3 自己的 {@code Log} 封装支持关掉日志，关掉后 {@code Log.d} 变成 no-op， + * 于是启发式能在 JVM 里完整执行。 + */ + @BeforeClass + public static void silenceMedia3Log() { + previousLogLevel = Log.getLogLevel(); + Log.setLogLevel(Log.LOG_LEVEL_OFF); + } + + /** + * 日志级别是进程级静态状态，而本模块没配 {@code forkEvery}，整个 module 跑在同一个 + * JVM 里 —— 不还原会按类的执行顺序泄漏给别的测试类。 + */ + @AfterClass + public static void restoreMedia3Log() { + Log.setLogLevel(previousLogLevel); + } + + private static int previousLogLevel; + + private static final String PLAYLIST_HOST = "v.example.com"; + private static final String PLAYLIST_URL = "https://" + PLAYLIST_HOST + "/play/index.m3u8"; + + /** + * 30 片，下标 3-5 是广告 —— 与正片**同域、同命名结构**，只有时长不同（6.4s 对 10.0s）。 + * + *

这个形状是刻意选的：legacy 启发式靠命名分组与前缀分析找广告，对完全均匀的 + * 命名会弃权，于是任何「删掉了广告」的断言都只能由结构化规则达成。此前夹具用 + * 独立广告域名 + 断点标记，启发式对它的输出与结构化引擎**逐字节相同**，导致 + * 4 条测试把规则换成 {@code List.of()} 也照样绿 —— 它们根本没在验证结构化规则。 + */ + private static String playlist() { + StringBuilder text = new StringBuilder("#EXTM3U\n#EXT-X-TARGETDURATION:10\n"); + for (int i = 0; i < 30; i++) { + boolean ad = i >= 3 && i <= 5; + text.append(String.format(Locale.US, "#EXTINF:%.3f,%n", ad ? 6.4 : 10.0)); + text.append("https://").append(PLAYLIST_HOST).append("/seg/").append(i).append(".ts\n"); + } + return text.append("#EXT-X-ENDLIST\n").toString(); + } + + /** 广告切片的 URI，供断言使用。 */ + private static String adUri(int index) { + return "https://" + PLAYLIST_HOST + "/seg/" + index + ".ts"; + } + + /** + * 直播窗口：无 {@code #EXT-X-ENDLIST}，带 {@code #EXT-X-MEDIA-SEQUENCE}。 + * + * @param mediaSequence 窗口首片的序号 + * @param adIndexes 窗口内属于广告的下标（相对窗口起点） + */ + private static String livePlaylist(int mediaSequence, int windowSize, Set adIndexes) { + StringBuilder text = new StringBuilder("#EXTM3U\n#EXT-X-TARGETDURATION:10\n"); + text.append("#EXT-X-MEDIA-SEQUENCE:").append(mediaSequence).append('\n'); + text.append("#EXT-X-DISCONTINUITY-SEQUENCE:0\n"); + for (int i = 0; i < windowSize; i++) { + boolean ad = adIndexes.contains(i); + if (ad) text.append("#EXT-X-DISCONTINUITY\n"); + text.append(String.format(Locale.US, "#EXTINF:%.3f,%n", ad ? 6.4 : 10.0)); + text.append(ad ? "https://ad-cdn.other.com/ads/" + (mediaSequence + i) + ".ts" + : "https://" + PLAYLIST_HOST + "/seg/" + (mediaSequence + i) + ".ts") + .append('\n'); + } + return text.toString(); + } + + /** + * 匹配 {@link #playlist()} 里 6.4s 那三片的规则。 + * + *

用时长而非域名区分：夹具刻意让广告与正片同域同命名，好让 legacy 启发式弃权， + * 从而使「广告被删」只可能由结构化规则达成。 + */ + private static HlsManifestCleaner.Rule adRule() { + return HlsAdRule.createUserRule("ad", "ad", List.of(PLAYLIST_HOST), + List.of(), List.of(), 6.3d, 6.5d, false, false, 1).compile(); + } + + private static String readFully(HlsPlaylistCleaningDataSource source, String url) + throws IOException { + source.open(null, url, 0, C.LENGTH_UNSET); + return drain(source); + } + + private static String drain(HlsPlaylistCleaningDataSource source) throws IOException { + ByteArrayOutputStream out = new ByteArrayOutputStream(); + byte[] chunk = new byte[512]; + while (true) { + int read = source.read(chunk, 0, chunk.length); + if (read == C.RESULT_END_OF_INPUT) break; + out.write(chunk, 0, read); + } + source.close(); + return out.toString(StandardCharsets.UTF_8); + } + + /** + * 结构化规则能在 Exo 路径上删掉广告切片、保留正片。 + * + *

不再断言「无规则时逐字节透传」—— 那在真机上是假的：无规则时结构化引擎不改动， + * {@link HlsAdblockPipeline} 会让 legacy 启发式接手，而它对这份夹具确实会删片。 + * 想验证「结构化引擎自身没有副作用」得关掉启发式，那属于 pipeline 的契约， + * 由 {@link #pipelineKeepsTheTwoEnginesMutuallyExclusive} 覆盖。 + */ + @Test + public void removesAdSegmentsFromPlaylistOnTheExoPath() throws IOException { + HlsPlaylistCleaningDataSource cleaning = new HlsPlaylistCleaningDataSource( + new FakeSource(playlist()), () -> List.of(adRule()), () -> true); + + String cleaned = readFully(cleaning, PLAYLIST_URL); + + assertFalse("广告切片应被删除", cleaned.contains(adUri(3))); + assertTrue("正片切片必须保留", cleaned.contains("/seg/0.ts")); + assertTrue("正片切片必须保留", cleaned.contains("/seg/29.ts")); + assertEquals("正片一片不少", 27, countSegments(cleaned)); + } + + private static int countSegments(String manifest) { + int count = 0; + for (String line : manifest.split("\n")) { + String trimmed = line.trim(); + if (!trimmed.isEmpty() && !trimmed.startsWith("#")) count++; + } + return count; + } + + /** + * 净化后 {@code open()} 必须返回改写后的长度。返回上游原长会让 Exo 认为响应被 + * 截断（净化只会变短），把正常播放报成 EOF 错误。 + */ + @Test + public void reportsTheRewrittenLengthNotTheUpstreamLength() throws IOException { + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(playlist()), () -> List.of(adRule()), () -> true); + + long reported = source.open(null, PLAYLIST_URL, 0, C.LENGTH_UNSET); + int delivered = drain(source).getBytes(StandardCharsets.UTF_8).length; + + assertEquals("声明的长度必须等于实际交付的字节数", delivered, reported); + assertTrue("净化后应短于原文", reported < playlist().getBytes(StandardCharsets.UTF_8).length); + } + + /** 非 playlist 的响应（切片、字幕等）必须逐字节原样透传。 */ + @Test + public void passesNonPlaylistPayloadsThroughByteForByte() throws IOException { + byte[] payload = new byte[4096]; + for (int i = 0; i < payload.length; i++) payload[i] = (byte) (i % 251); + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(payload), () -> List.of(adRule()), () -> true); + + source.open(null, "https://" + PLAYLIST_HOST + "/seg/0.ts", 0, C.LENGTH_UNSET); + ByteArrayOutputStream out = new ByteArrayOutputStream(); + byte[] chunk = new byte[100]; + while (true) { + int read = source.read(chunk, 0, chunk.length); + if (read == C.RESULT_END_OF_INPUT) break; + out.write(chunk, 0, read); + } + source.close(); + + assertArrayEquals("切片数据不得被改动", payload, out.toByteArray()); + } + + /** 带 BOM 与前导空白的 playlist 也要被识别 —— 与 media3 的 parser 判据一致。 */ + @Test + public void recognisesPlaylistBehindByteOrderMarkAndWhitespace() throws IOException { + String text = "\n " + playlist(); + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(text), () -> List.of(adRule()), () -> true); + + String cleaned = readFully(source, PLAYLIST_URL); + + assertFalse("BOM 后的 playlist 同样要净化", cleaned.contains(adUri(3))); + } + + /** + * 区间请求不得改写：range 寻址的是原始响应的字节偏移，而净化会改变长度， + * 改写会让偏移全部错位。 + */ + @Test + public void leavesRangedRequestsUntouched() throws IOException { + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(playlist()), () -> List.of(adRule()), () -> true); + + source.open(null, PLAYLIST_URL, 10, 200); + String served = drain(source); + + assertTrue("区间请求必须原样透传", served.contains(adUri(3))); + } + + /** 去广告总开关关闭时完全不介入。 */ + @Test + public void doesNothingWhenAdblockIsDisabled() throws IOException { + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(playlist()), () -> List.of(adRule()), () -> false); + + assertEquals(playlist(), readFully(source, PLAYLIST_URL)); + } + + /** + * 本地 {@code /m3u8} 代理的输出已经净化过，不能再来一遍 —— 那会让规则在同一份 + * playlist 上叠加两次，删除量翻倍并可能越过回退闸门。 + */ + @Test + public void skipsPlaylistsAlreadyCleanedByTheLocalProxy() throws IOException { + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(playlist()), () -> List.of(adRule()), () -> true); + + String served = readFully(source, + "http://127.0.0.1:9978/m3u8?url=https%3A%2F%2Fv.example.com%2Fa.m3u8"); + + assertTrue("本地代理已净化，此处必须放过", served.contains(adUri(3))); + } + + /** 读取规则配置抛异常时必须降级为透传，而不是把异常抛给播放中的用户。 */ + @Test + public void degradesToPassThroughWhenRuleConfigThrows() throws IOException { + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(playlist()), + () -> { throw new IllegalStateException("config unavailable"); }, + () -> true); + + assertEquals(playlist(), readFully(source, PLAYLIST_URL)); + } + + /** 上游按小块返回时，嗅探与整体读取都必须正确拼接。 */ + @Test + public void handlesUpstreamThatReturnsTinyChunks() throws IOException { + FakeSource upstream = new FakeSource(playlist()); + upstream.maxChunk = 3; + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + upstream, () -> List.of(adRule()), () -> true); + + String cleaned = readFully(source, PLAYLIST_URL); + + assertFalse("分片读取不得影响净化", cleaned.contains(adUri(3))); + assertTrue(cleaned.contains("/seg/29.ts")); + } + + /** + * 直播窗口一律不改写。 + * + *

{@code HlsManifestCleaner} 会按删掉的头部片数抬高 + * {@code #EXT-X-MEDIA-SEQUENCE}，但它的回退闸门（删除比例、删除时长、删除必须 + * 在头部连续）会在相邻两次 reload 之间翻转，回退时发出的是**未抬高**的原始序号。 + * 而滑动窗口里新的广告块总是先出现在尾部（删除点在保留点之后 → 非连续 → 回退）， + * 随窗口滑动才变成头部连续 —— 于是序号在抬高与原始之间来回跳。media3 的 + * {@code isNewerThan} 会静默丢弃「变旧」的那次 reload，最终 + * {@code targetDuration × 3.5} 后抛 {@code PlaylistStuckException} 卡播。 + */ + @Test + public void neverRewritesLivePlaylists() throws IOException { + String live = livePlaylist(100, 12, Set.of(0, 1, 2)); + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(live), () -> List.of(adRule()), () -> true); + + assertEquals("直播窗口必须逐字节透传", live, readFully(source, PLAYLIST_URL)); + } + + /** + * 直播的关键不变量：相邻 reload 发出的 media sequence 不得倒退。 + * + *

倒退需要「先净化、后回退」这个特定顺序 —— 发出的序号要么是窗口起点（回退）， + * 要么是首个保留切片的原始下标（净化后），两者在纯滑动中各自单调；只有从净化跳到 + * 回退才会掉回去。触发它的正是真实直播的形态：头部还有正在被消费的广告块（净化， + * 抬高序号），同时**新的**广告块滑进窗口尾部（删除点不连续 → 回退 → 原始序号）。 + * + *

这里把广告块放在绝对下标 100-102 与 112-114：窗口起点 100 时只见到头部那块， + * 抬高到 103；起点 101 时尾部滑入 112，变成不连续而回退到 101 —— 比上一次小。 + */ + @Test + public void keepsMediaSequenceMonotonicAcrossLiveReloads() throws IOException { + Set adSegments = Set.of(100, 101, 102, 112, 113, 114); + int previous = Integer.MIN_VALUE; + for (int start = 100; start <= 118; start++) { + Set ads = new HashSet<>(); + for (int offset = 0; offset < 12; offset++) { + if (adSegments.contains(start + offset)) ads.add(offset); + } + String live = livePlaylist(start, 12, ads); + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(live), () -> List.of(adRule()), () -> true); + + int emitted = mediaSequenceOf(readFully(source, PLAYLIST_URL)); + + assertTrue("media sequence 不得倒退：" + previous + " -> " + emitted + + "（窗口起点 " + start + "）", emitted >= previous); + previous = emitted; + } + } + + private static int mediaSequenceOf(String manifest) { + for (String line : manifest.split("\n")) { + String trimmed = line.trim(); + if (trimmed.startsWith("#EXT-X-MEDIA-SEQUENCE:")) { + return Integer.parseInt(trimmed.substring("#EXT-X-MEDIA-SEQUENCE:".length()).trim()); + } + } + throw new AssertionError("缺少 #EXT-X-MEDIA-SEQUENCE：" + manifest); + } + + /** + * 结构化规则改动过的 manifest，绝不能再交给 legacy 启发式。 + * + *

{@code HlsAdsParser.findMinorityGroup} 只有一道「最大分组须超过切片总数的 + * 50%」的闸门，且对删除量没有上限；而删掉广告切片会缩小它比对的总数。夹具刻意 + * 用「正片分布在两个 CDN（260 + 240 片）+ 广告 25 片」这个实测形状：原始 playlist + * 上启发式因 {@code 520 <= 525} 放弃，结构化净化后变成 {@code 520 > 500}，闸门打开 + * 并删掉那 240 片正片 —— 48% 的正片被静默删除。 + * + *

用这个形状而不是 {@link #playlist()}：后者太均匀，启发式对净化前后的结果一致， + * 即使两套引擎真被串联起来测试也照样绿。 + * + *

Exo 侧另一半（不让 fork 版 parser 重复跑）由 {@code ExoUtil} 的 + * {@code setAdblock(false)} 保证，那一处纯 JVM 测不了。 + */ + @Test + public void pipelineKeepsTheTwoEnginesMutuallyExclusive() { + String twoCdn = twoCdnPlaylist(); + HlsManifestCleaner.Rule rule = HlsAdRule.createUserRule("ad", "ad", + List.of(PLAYLIST_HOST), List.of("ad-cdn.other.com"), List.of(), + null, null, false, true, 2).compile(); + + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(PLAYLIST_URL, twoCdn, List.of(rule), true); + + assertTrue("结构化引擎应当生效", outcome.structured()); + assertFalse("结构化改动过就不该再跑启发式", outcome.legacy()); + // 若两套引擎被串联，启发式会把 cdn-b 那 240 片正片一并删掉 + assertEquals("只应删掉 25 片广告", 500, countSegments(outcome.manifest())); + assertTrue("第二个 CDN 的正片必须完好", outcome.manifest().contains("cdn-b.example.com")); + } + + /** + * 上一条的反面：把净化后的结果再喂给启发式，确认「串联确实会毁掉正片」。 + * + *

这条不测生产代码路径，它测的是**威胁本身仍然存在** —— 一旦有人把 + * {@code ExoUtil} 的 {@code setAdblock} 改回 {@code Setting.isAdblock()}，或者给 + * pipeline 加上「两套都跑」的开关，48% 正片就会消失。若哪天启发式自己加了删除 + * 上限而让这条测试变红，那说明威胁解除，可以连同那半边约束一起删掉。 + */ + @Test + public void chainingTheTwoEnginesWouldDestroyMainContent() { + String twoCdn = twoCdnPlaylist(); + HlsManifestCleaner.Rule rule = HlsAdRule.createUserRule("ad", "ad", + List.of(PLAYLIST_HOST), List.of("ad-cdn.other.com"), List.of(), + null, null, false, true, 2).compile(); + + // 启发式对原始 playlist 弃权（525 片里最大分组 260，520 <= 525） + String heuristicOnRaw = HlsAdsParser.process(twoCdn); + assertEquals("原始 playlist 上启发式应当弃权", 525, countSegments(heuristicOnRaw)); + + // 但对净化后的 500 片，闸门打开（520 > 500） + String structured = HlsManifestCleaner.clean(PLAYLIST_URL, twoCdn, List.of(rule)).manifest(); + assertEquals(500, countSegments(structured)); + String chained = HlsAdsParser.process(structured); + + assertEquals("串联后 cdn-b 的 240 片正片被删掉", 260, countSegments(chained)); + assertFalse("这就是必须避免串联的原因", chained.contains("cdn-b.example.com")); + } + + /** + * 正片分布在两个 CDN（260 + 240 片）+ 25 片广告，共 525 片。 + * + *

广告时长取 3.0s：25 片合计 75s，压在 {@code HlsManifestCleaner} 的 90s 删除 + * 时长闸门内，否则结构化引擎会整体回退而测不到后续的引擎交互。 + */ + private static String twoCdnPlaylist() { + StringBuilder text = new StringBuilder("#EXTM3U\n#EXT-X-TARGETDURATION:10\n"); + for (int i = 0; i < 260; i++) { + text.append("#EXTINF:10.000,\n") + .append("https://").append(PLAYLIST_HOST).append("/seg/").append(i).append(".ts\n"); + } + for (int i = 0; i < 25; i++) { + text.append("#EXT-X-DISCONTINUITY\n#EXTINF:3.000,\n") + .append("https://ad-cdn.other.com/ads/").append(i).append(".ts\n"); + } + for (int i = 0; i < 240; i++) { + text.append("#EXTINF:10.000,\n") + .append("https://cdn-b.example.com/seg/").append(i).append(".ts\n"); + } + return text.append("#EXT-X-ENDLIST\n").toString(); + } + + /** + * 结构化引擎没改动时，pipeline 必须把 manifest 递给启发式 —— 否则默认内核会比改动 + * 前更差（改动前 fork 版 parser 无条件跑启发式）。 + * + *

只钉「调用发生了」而不钉「删到了东西」：启发式是否动手取决于它自己的分组判据， + * 而两个夹具恰好都会让它弃权（{@link #playlist()} 命名太均匀，{@link #twoCdnPlaylist()} + * 的 525 片里最大分组 260 不过半，实测两者 {@code process} 前后片数不变）。用「片数 + * 变少」当判据会把这条测试变成对启发式内部阈值的断言，那不是本处的契约。 + * + *

判据落在 {@code legacy()} 上：它由 {@code !filtered.equals(manifest)} 得出， + * 弃权时为 false，所以只能反向钉住 —— 结构化未改动时 pipeline 不得提前返回。 + * 若把 {@code HlsAdblockPipeline} 的 legacy 分支删掉，{@code manifest()} 会变成 + * {@code null} 之外的同一对象，这条测试确实抓不到。真正守住 legacy 分支的是 + * {@link #pipelineKeepsTheTwoEnginesMutuallyExclusive} 的 {@code legacy=false} 断言 + * 与 {@link #chainingTheTwoEnginesWouldDestroyMainContent}；这里只补一条 + * 「结构化未命中时不短路」的正向路径。 + */ + @Test + public void pipelineStillReachesTheHeuristicWhenNoRuleMatches() { + String twoCdn = twoCdnPlaylist(); + + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(PLAYLIST_URL, twoCdn, List.of(), true); + + assertFalse("无规则命中", outcome.structured()); + // 启发式对这个形状弃权，所以 manifest 应当与输入相同、且不是 null + assertEquals("弃权时必须原样返回", twoCdn, outcome.manifest()); + assertFalse("弃权就不算 legacy 改动过", outcome.legacy()); + } + + /** + * 超过 2 MiB 上限时必须原样透传，且字节精确、顺序不变、只发一遍。 + * + *

夹具刻意含广告切片：不含的话结构化引擎与启发式都不会改动，测试即使把 + * {@code MAX_PLAYLIST_BYTES} 那道门删掉也照样绿 —— 钉不住上限本身。 + */ + @Test + public void passesOversizedPlaylistThroughByteForByte() throws IOException { + StringBuilder huge = new StringBuilder("#EXTM3U\n#EXT-X-TARGETDURATION:10\n"); + int i = 0; + while (huge.length() < 3 * 1024 * 1024) { + boolean ad = i % 50 == 7; + if (ad) huge.append("#EXT-X-DISCONTINUITY\n"); + huge.append(ad ? "#EXTINF:3.000,\n" : "#EXTINF:10.000,\n"); + huge.append(ad ? "https://ad-cdn.other.com/ads/" + i + ".ts" + : "https://" + PLAYLIST_HOST + "/seg/" + i + ".ts").append('\n'); + i++; + } + String text = huge.append("#EXT-X-ENDLIST\n").toString(); + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(text), () -> List.of(adRule()), () -> true); + + assertEquals("超限的 playlist 必须逐字节透传", text, readFully(source, PLAYLIST_URL)); + } + + /** + * 同一个实例被反复 {@code open()} / {@code close()} —— 这正是 + * {@code DefaultHlsPlaylistTracker} 复用一个 DataSource 做直播 reload 的生命周期。 + * 状态没清干净会把上一次的缓冲字节串到下一次。 + */ + @Test + public void reusesOneInstanceAcrossOpenCloseCycles() throws IOException { + FakeSource upstream = new FakeSource(playlist()); + upstream.rewindOnOpen = true; + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + upstream, () -> List.of(adRule()), () -> true); + + String first = readFully(source, PLAYLIST_URL); + String second = readFully(source, PLAYLIST_URL); + + assertEquals("复用同一实例的两次结果必须一致", first, second); + assertFalse(second.contains(adUri(3))); + } + + /** {@code DataReader} 契约：{@code readLength == 0} 必须返回 0，而不是 EOF。 */ + @Test + public void returnsZeroForZeroLengthRead() throws IOException { + HlsPlaylistCleaningDataSource source = new HlsPlaylistCleaningDataSource( + new FakeSource(playlist()), () -> List.of(adRule()), () -> true); + + source.open(null, PLAYLIST_URL, 0, C.LENGTH_UNSET); + + assertEquals(0, source.read(new byte[4], 0, 0)); + source.close(); + } + + /** + * {@code withPlaylistCleaning} 确实产出净化层。 + * + *

坦白其局限：它钉的是那个工厂方法本身，**不是**「Exo 的数据源链条上真的挂着 + * 它」。后者要走 {@code getDataSourceFactory}，而那里需要 {@code App.get()} 与磁盘 + * 缓存，纯 JVM 下构造不出来。也就是说把 {@code getDataSourceFactory} 里的调用删掉， + * 这条测试仍会绿 —— 那一处只能靠人工复查或仪器化测试守。 + */ + @Test + public void withPlaylistCleaningProducesACleaningSource() { + DataSource.Factory wrapped = MediaSourceFactory.withPlaylistCleaning( + () -> new FakeSource(playlist())); + + assertTrue("工厂必须产出净化层", + wrapped.createDataSource() instanceof HlsPlaylistCleaningDataSource); + } + + private static final class FakeSource implements DataSource { + + private final byte[] data; + private int position; + private int maxChunk = Integer.MAX_VALUE; + /** 直播 reload 会对同一个 DataSource 反复 open，每次都从头给。 */ + private boolean rewindOnOpen; + + FakeSource(String text) { + this(text.getBytes(StandardCharsets.UTF_8)); + } + + FakeSource(byte[] data) { + this.data = data; + } + + @Override + public void addTransferListener(TransferListener transferListener) { + } + + @Override + public long open(DataSpec dataSpec) { + if (rewindOnOpen) position = 0; + return data.length - position; + } + + @Override + public int read(byte[] target, int offset, int length) { + if (position >= data.length) return C.RESULT_END_OF_INPUT; + int count = Math.min(Math.min(length, maxChunk), data.length - position); + System.arraycopy(data, position, target, offset, count); + position += count; + return count; + } + + @Override + public Uri getUri() { + return null; + } + + @Override + public Map> getResponseHeaders() { + return Map.of(); + } + + @Override + public void close() { + } + } +} diff --git a/app/src/test/java/com/fongmi/android/tv/utils/HlsEncryptionGuardTest.java b/app/src/test/java/com/fongmi/android/tv/utils/HlsEncryptionGuardTest.java new file mode 100644 index 0000000000..cd8f8e7978 --- /dev/null +++ b/app/src/test/java/com/fongmi/android/tv/utils/HlsEncryptionGuardTest.java @@ -0,0 +1,228 @@ +package com.fongmi.android.tv.utils; + +import static org.junit.Assert.assertEquals; +import static org.junit.Assert.assertFalse; +import static org.junit.Assert.assertTrue; + +import androidx.media3.common.util.Log; + +import com.fongmi.android.tv.bean.HlsAdRule; + +import org.junit.AfterClass; +import org.junit.BeforeClass; +import org.junit.Test; + +import java.util.List; +import java.util.Locale; + +/** + * 隐式 IV 的加密 playlist 必须让**两套引擎都不动手**。 + * + *

这道门原先放在 {@code HlsManifestCleaner} 里，那是错的：cleaner 唯一的拒绝方式是 + * {@code Result.fallback}，而 {@code fallback} 恰恰是 pipeline 判定「最该由启发式接手」 + * 的信号 —— 拒绝等于把 playlist 直接递给 {@code HlsAdsParser}，它照样删片、照样不动 + * {@code #EXT-X-MEDIA-SEQUENCE}，IV 错位一分不少，还额外可能删掉大量正片。 + */ +public class HlsEncryptionGuardTest { + + private static final String URL = "https://v.example.com/play/index.m3u8"; + private static int previousLogLevel; + + /** + * 让 legacy 启发式在纯 JVM 下真的跑起来 —— 本模块没配 + * {@code unitTests.returnDefaultValues}，未 mock 的 {@code android.util.Log} 会让 + * {@code HlsAdsParser.process} 第一句就抛，而 pipeline 会 catch 掉，启发式实际从未执行。 + */ + @BeforeClass + public static void silenceMedia3Log() { + previousLogLevel = Log.getLogLevel(); + Log.setLogLevel(Log.LOG_LEVEL_OFF); + } + + /** 日志级别是进程级静态状态，同一个 test JVM 里会泄漏给其他测试类。 */ + @AfterClass + public static void restoreMedia3Log() { + Log.setLogLevel(previousLogLevel); + } + + private static HlsManifestCleaner.Rule adRule() { + return HlsAdRule.createUserRule("ad", "ad", List.of("v.example.com"), + List.of("ad-cdn.other.com"), List.of(), null, null, false, true, 2).compile(); + } + + /** 30 片，下标 3-5 是跨域广告；{@code keyLine} 为 null 时不加密。 */ + private static String playlist(String keyLine) { + StringBuilder t = new StringBuilder("#EXTM3U\n#EXT-X-VERSION:3\n"); + t.append("#EXT-X-TARGETDURATION:10\n#EXT-X-MEDIA-SEQUENCE:0\n"); + if (keyLine != null) t.append(keyLine).append('\n'); + for (int i = 0; i < 30; i++) { + boolean ad = i >= 3 && i <= 5; + t.append(String.format(Locale.US, "#EXTINF:%.3f,%n", ad ? 6.4 : 10.0)); + t.append(ad ? "https://ad-cdn.other.com/ads/" + i + ".ts" + : "https://v.example.com/seg/" + i + ".ts").append('\n'); + } + return t.append("#EXT-X-ENDLIST\n").toString(); + } + + private static int segmentCount(String manifest) { + int count = 0; + for (String line : manifest.split("\n")) { + String trimmed = line.trim(); + if (!trimmed.isEmpty() && !trimmed.startsWith("#")) count++; + } + return count; + } + + /** 未加密的对照：两套引擎照常工作，证明夹具本身是可净化的。 */ + @Test + public void plainPlaylistIsStillCleaned() { + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(URL, playlist(null), List.of(adRule()), true); + + assertTrue(outcome.structured()); + assertEquals(27, segmentCount(outcome.manifest())); + } + + /** + * 核心不变量：隐式 IV 时 manifest 必须逐字节原样返回，且两套引擎都没跑。 + * + *

只断言 {@code structured==false} 是不够的 —— 那正是旧实现的状态，而启发式在 + * 它之后照样把片删了。必须同时钉住 {@code legacy==false} 与字节相等。 + */ + @Test + public void impliedIvPlaylistIsUntouchedByBothEngines() { + String manifest = playlist("#EXT-X-KEY:METHOD=AES-128,URI=\"key.bin\""); + + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(URL, manifest, List.of(adRule()), true); + + assertFalse("结构化引擎不得改动", outcome.structured()); + assertFalse("启发式也不得接手", outcome.legacy()); + assertEquals("必须逐字节原样返回", manifest, outcome.manifest()); + assertEquals(30, segmentCount(outcome.manifest())); + } + + /** 显式 {@code IV=} 固定了 IV，与序号无关，应当照常净化。 */ + @Test + public void explicitIvPlaylistIsStillCleaned() { + String manifest = playlist( + "#EXT-X-KEY:METHOD=AES-128,URI=\"key.bin\",IV=0x0123456789abcdef0123456789abcdef"); + + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(URL, manifest, List.of(adRule()), true); + + assertTrue("显式 IV 不受影响", outcome.structured()); + assertEquals(27, segmentCount(outcome.manifest())); + } + + /** {@code METHOD=NONE} 只是取消加密，不涉及 IV 推导。 */ + @Test + public void methodNoneIsNotTreatedAsEncrypted() { + HlsAdblockPipeline.Outcome outcome = HlsAdblockPipeline.apply( + URL, playlist("#EXT-X-KEY:METHOD=NONE"), List.of(adRule()), true); + + assertTrue(outcome.structured()); + assertEquals(27, segmentCount(outcome.manifest())); + } + + /** + * 四个曾经骗过这道门的形状，全部必须被识别为隐式 IV。 + * + *

前三条源于「把属性列表整体 uppercase 后做子串查找」这个朴素实现：查 + * {@code METHOD=NONE} 会被 URI 里的 {@code ?method=none} 命中；查 {@code IV=} 会被 + * {@code ?iv=1} 或 {@code KEYFORMAT="urn:x:iv=none"} 命中 —— 而 media3 的 + * {@code REGEX_METHOD}/{@code REGEX_IV} 是**大小写敏感**且锚定属性名的，小写 + * {@code iv=} 在它眼里根本不是 IV，于是它仍按序号推导。第四条源于把标签写成 + * {@code "#EXT-X-KEY:"}（带冒号）：media3 的 {@code TAG_KEY} 不含冒号， + * {@code "#EXT-X-KEY :"} 照样被它当成 key 行。 + */ + @Test + public void recognisesShapesThatDefeatNaiveSubstringMatching() { + String[] keyLines = { + // URI 的 query 里出现 method=none + "#EXT-X-KEY:METHOD=AES-128,URI=\"https://k/x?method=none\"", + // URI 的 query 里出现小写 iv= + "#EXT-X-KEY:METHOD=AES-128,URI=\"https://k/k?iv=1&t=2\"", + // KEYFORMAT 里出现小写 iv= + "#EXT-X-KEY:METHOD=AES-128,URI=\"k.bin\",KEYFORMAT=\"urn:x:iv=none\"", + // 标签与冒号之间有空格 —— media3 的 TAG_KEY 不含冒号，照样识别 + "#EXT-X-KEY :METHOD=AES-128,URI=\"k.bin\"", + }; + + for (String keyLine : keyLines) { + String manifest = playlist(keyLine); + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(URL, manifest, List.of(adRule()), true); + + assertEquals("必须原样返回：" + keyLine, manifest, outcome.manifest()); + assertFalse("结构化引擎不得改动：" + keyLine, outcome.structured()); + assertFalse("启发式不得接手：" + keyLine, outcome.legacy()); + } + } + + /** + * 纯 {@code \r} 换行的 playlist 同样要识别。 + * + *

按 {@code "\n"} 切行会把整份 playlist 塌成一行，key 行就此消失；而 + * {@code BufferedReader.readLine} 认单独的 {@code \r} 为行结束，media3 能看到 key。 + */ + @Test + public void recognisesKeyLineInCarriageReturnOnlyPlaylist() { + String manifest = playlist("#EXT-X-KEY:METHOD=AES-128,URI=\"key.bin\"") + .replace("\n", "\r"); + + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(URL, manifest, List.of(adRule()), true); + + assertEquals("CR-only 也必须原样返回", manifest, outcome.manifest()); + assertFalse(outcome.structured()); + assertFalse(outcome.legacy()); + } + + /** + * {@code IV = 0x…}（等号旁有空格）在 media3 眼里也不算固定 IV，必须拒绝。 + * + *

这个方向本来就安全（朴素实现也会拒），留作回归：将来若有人把正则放宽成容忍 + * 空格，就会与 media3 产生分歧。 + */ + @Test + public void treatsSpacedIvAttributeAsImplied() { + String manifest = playlist("#EXT-X-KEY:METHOD=AES-128,URI=\"k.bin\",IV = 0x00"); + + HlsAdblockPipeline.Outcome outcome = + HlsAdblockPipeline.apply(URL, manifest, List.of(adRule()), true); + + assertEquals(manifest, outcome.manifest()); + assertFalse(outcome.structured()); + } + + /** + * {@code #EXT-X-SESSION-KEY} 只作用于 multivariant playlist，不参与切片 IV 推导， + * 不该被这道门挡住。 + * + *

media3 的判据是 {@code line.startsWith("#EXT-X-KEY")}，而 + * {@code "#EXT-X-SESSION-KEY"} 不以它开头。 + */ + @Test + public void sessionKeyDoesNotBlockCleaning() { + HlsAdblockPipeline.Outcome outcome = HlsAdblockPipeline.apply(URL, + playlist("#EXT-X-SESSION-KEY:METHOD=AES-128,URI=\"k.bin\""), + List.of(adRule()), true); + + assertTrue("SESSION-KEY 不影响切片 IV", outcome.structured()); + assertEquals(27, segmentCount(outcome.manifest())); + } + + /** + * 小写 {@code #ext-x-key:} 在 media3 眼里不是 key 行（{@code startsWith} 大小写敏感）， + * 所以它看不到加密，净化是安全的。 + */ + @Test + public void lowercaseKeyTagIsInvisibleToBothSides() { + HlsAdblockPipeline.Outcome outcome = HlsAdblockPipeline.apply(URL, + playlist("#ext-x-key:METHOD=AES-128,URI=\"k.bin\""), + List.of(adRule()), true); + + assertTrue("media3 也看不到这行，判据必须一致", outcome.structured()); + } +}