fix(errors): stop the fingerprint shattering on any hex run of an odd length

`hexDigest` recognised only 64/40/32/24 (sha256/sha1/md5/uuid). Every other hex
run fell through to `integerFb`, which emitted `{integer}` for the leading digits
and rescanned the letters — so `SB-3FDBA2D0A0A2` and `SB-90770E7BA9FD` normalised
to DIFFERENT strings. Every order id became its own error group, its own issue
row, and its own notification, which is the "we keep getting told about errors
we have seen before" report.

Measured over 30 days of prod error_patterns: 1486 distinct hashes where 783 is
right. One group alone — the same `insert into orders` failure — was 178.

12 as the floor, not 8: at 8 this starts eating short commit SHAs and hex-looking
words that are the only thing telling two errors apart, and the corpus says that
reach is worth 13 groups (770 vs 783). 16 leaves the 12-15 char ids that dominate
the fan-out (1134).

Over-merge checked against the corpus, not assumed: the groups this newly merges
are one bug each — `SB-{hex} not processing` x117, a single TLS handshake failure
x65 — while genuinely different conditions (`already cancelled` vs `already
completed`) stay apart.

This changes fingerprints, so every currently-open pattern mints a new hash on its
next occurrence: expect one final duplication wave as the corrected groups form.
This commit is contained in:
Anthony Alaribe 2026-09-03 09:52:27 +02:00
commit e00e7ec78d