test_recall real-query floor leg couples to self-index corpus stats — large refactors false-red it (2 sightings) #420

Defect (test-stability, systemic — 2 sightings today)

test_recall's "real query: top rows clear the floor" leg (tests/test_recall.py:500-502) asserts that recall.search("graph signal wiring edges")[:3] keeps at least one un-weak row against the suite's SELF-INDEX of the checkout. The weak mark is floor-gated (cos < 0.48 AND bm25 < 6.0, issue #297), so the leg silently couples to the repository's own BM25F corpus statistics: any large textual refactor shifts IDF/length norms until all top-3 rows fall under the floor — with ZERO behavior change in the engine.

Sightings (2026-09-20/21): PR #416's merge-commit CI red (decode: corpus-delta into pre-#407 main; rebase + fresh CI green); PR #419's deterministic red (+146 common.py / −86 c / −97 cpp shifts stats; passes on pristine main, fails with diff, fails with the tooth test removed — pure corpus effect). Both required lane-hours to decode.

Fix shape

De-couple the leg from the mutable self-index, one of:

• rank-shape assertion: assert the garbage-query rows are ALL weak while the real query's top-k contains the INTENDED target file (membership, not floor-clearing) — floor-clearing already pinned by the constants check + bench/golden; or
• stable fixture corpus for §11 (the recall/explore: no absolute relevance floor (garbage queries get confident hits); explore blames the backend for an empty index #297 calibration intent, frozen inputs).

Teeth

Mutation: corpus-only textual churn (comment moves) must NOT flip the leg; a genuine floor regression (floor raised above real-query evidence) must still fail it.

Context

Filed by GK during the #419 gate (ruling: query recalibrated in-PR as explicit pinned delta; floors are the pinned contract — bench/golden + the constants check). This issue owns the systemic de-coupling so the NEXT big refactor doesn't pay the decode again.