fix(endpoints): let the gates below the LLM verdict actually pass

The group review was already returning the right answer — "param", shape
"dot-separated metric name", confirmed three times for the family behind the
reported alert. Four gates under it threw that away, and only 3 of 60 param
verdicts fleet-wide had ever been applied.

- mergeEvidenceMet required the group to grow between passes. A group is only
  re-asked when its membership changes, so a settled project never reaches the
  second confirmation: 55 of 60 verdicts died here, and 983 of 1,069 reviews sit
  at one confirmation. Growth becomes "has not shrunk", and param groups are
  offered a re-ask after seven days so the second pass can happen at all.
- The route-word veto was all-or-nothing, and whole naming schemes read as
  words: one app_recommendations_counter vetoed a family of 24. It is now a
  fraction, sitting in the gap the fleet shows (param p90 0.33, routes p10 0.40).
- shapeAgreementOk had no floor, so nine endpoints returning one identical shape
  scored 1 > 0.9 and were refused. The gate meant to catch disagreement was
  rejecting the most agreeable groups there are.
- Shape agreement was pooled across a whole group rather than measured per
  template the merge produces, charging a mid-path verdict for the difference
  between .../{param}/, .../{param}/exemplars and .../{param}/breakdown.

Notifications, independently: the fast path sent any label containing "{",
reading "one segment is templated" as "cannot collapse further" — so
/p/{uuid}/metrics/details/<metric>/exemplars skipped the discovery deferral on
the strength of a {uuid} five segments away. That path is gone, and siblings
that differ at one position now render as one line rather than nine.

Projected over the fleet: 13 groups, 1,754 endpoint rows, capped at three groups
per run with the existing quarantine as the backstop.
This commit is contained in:
Anthony Alaribe 2026-09-01 22:25:46 +02:00
commit b265edd9ef
3 changed files with 252 additions and 66 deletions