Named for what ties this release together: three separate things that looked like they were working and were not. A BC fallback that had never once run, a validation gate that passed on sampling luck, and a baseline file that version control was silently discarding. None of them were failing loudly enough to notice.
TypeError
on every single request since the day it was written six months ago, with the failure
swallowed by a debug-level exception handler. It was broken in three independent places
at once, so fixing the reported defect would only have moved the failure one line
downstream. It was removed rather than revived: the component is an uncalibrated older
model reporting a hardcoded confidence, and it was wired to override the calibrated
ensemble precisely when that ensemble reported low confidence. Behaviour preservation
was measured across 2,560 caliber/weight/model/velocity cells, not assumed.*.json ignore rule from silently dropping the
gate baseline — the newly added baseline was untracked, so the gate would
have quietly degraded to its weaker ceiling-only form on any fresh clone while still
reporting a pass.cli_tokens and
monitoring_service — a transient Firestore blip on the
entitlement path denied paying customers, and surfaced as a 401, which is a 4xx and
therefore invisible to the API's 5xx alarm.The BC5D staleness gate had been red since August 14 from two AWS-migration commits that moved provenance files without changing any math. The behavioural probe confirms no correction cell moved, so the published v2.4.0 tables remain correct and no regeneration was needed; the baseline has been refreshed to re-arm the gate.