Keep recall-collapse baselines stable through silent days

[hamzamerzic]

Summary

• keep silent days from rewriting an established recall baseline during a sustained outage
• require three active historical days so one busy day cannot create a false baseline
• cover prolonged collapse and sparse-history behavior with regression tests

Why

PR #38 introduced the recall-collapse signal using a trailing seven-day median. During a multi-day outage, silent days eventually drove that median below the alert threshold, so the outage became its own normal and the alert cleared while recall remained inactive.

Ignoring silent days alone would make a single busy day sufficient evidence. This follow-up uses a bounded 14-day nonzero window and requires three active days before treating the median as established.

Tests

• PYTHONPATH=. pytest -q (71 tests, 2 subtests)

[hamzamerzic]

Möbius agent review

Result: Blocking queue finding: do not merge this alongside #43. PR #43 replaces the sliding automatic verdict with a durable completion checkpoint and ordered recall evidence, which is the coherent solution to the same failure. Close #42 as superseded by #43.

I reread the full diff and checked the current GitHub workflow result; the reviewed revision is green unless the queue note above calls for a later rebase.

This is a disclosed agent review posted by the PR author’s account, not an independent maintainer approval.