review-bench fixture maintenance: fold baseline-r1 adjudicated unseeded defects into truth/fixtures

[SSFSKIM]

Problem & intent

Baseline-r1 adjudication surfaced genuine unseeded defects and two imperfect baits in the X1 fixtures (tests/review-bench/). Before the next scored run-id (C2's re-run), fold the adjudicated reality back into the fixtures so recall/FP stay clean: either fix each unseeded defect in base/patch or promote it into truth.json; tighten the two baits that drew defensible flags. Full list: results/2026-07-26-baseline-r1/scores.json genuine_unseeded_excluded + notes.md.

Constraints

• Do not change any seeded bug or its truth entry (comparability across run-ids).
• The case3 fixture needs an explicit single-threaded statement (README/base comment) so the recurring TOCTOU finding is unambiguously refutable — both engines flagged it.
• case5's bounded-retention bait: state the retention bound where a reviewer will see it.

Success criteria

• A fresh engine run on the updated fixtures produces no finding that is neither truth-matched nor a deliberate bait.
• scores.json for the next run-id needs no "genuine-unseeded" escape hatch entries for known items.

Open questions

• none — mechanical fixture maintenance guided by the adjudication record.

Decision log

• Registered at C1 close per the deferral rule (spawned-by C1 spike: review methodology comparison + X1 benchmark + headless probe #28).

[SSFSKIM]

[board] ready-for-agent: before C2's bench re-run

[SSFSKIM]

[board] in-progress: delegated to in-session subagent under C2 (#29) work; fixture edits per baseline-r1 adjudication

[SSFSKIM]

[board] done: fixture maintenance landed in-session (subagent); acceptance's fresh-run check rides C2's r2 bench (next step in-session)