review-bench fixture maintenance round 2: fold C2-run adjudications + resolve u2 disagreement

[SSFSKIM]

Problem & intent

C2's scored runs (post-#36 fixtures) surfaced a second round of adjudicated genuine-unseeded defects plus one cross-engine disagreement on a promoted truth entry. Fold them into tests/review-bench/ truth/fixtures the same way #36 did, so the NEXT scored run-id needs no escape-hatch exclusions.

Items (details in results/2026-07-26-c2-scores.json new_unseeded_candidates_excluded and the per-run notes.md):

• case4 source-dir symlink probe — cross-engine convergent (codex r3, argus plain r2, argus high r2 [P1]); genuine, fix or promote
• case4 staging-filesystem free-space gap (argus high r2 [P2] PLAUSIBLE)
• case4 find -type f drops symlinked archive files (both argus runs [P3])
• case3 expected_revision: null skips validation (codex r3 [P2]); adjudicate then fix or promote
• case1 null config entry crashes load_tenant_limits (argus high r2 [P3])
• case1 explicit-null tenant stringified to "None" (argus high r2 [P3])
• case4-u2 disagreement: codex finds the remote-umask exposure every run; an argus high verifier actively REFUTED it in r2. Pull the refutation reasoning from the run transcript, decide whether u2 stands as truth (and whether the verifier text needs a nudge) — do not silently drop either side.

Constraints

• Same as review-bench fixture maintenance: fold baseline-r1 adjudicated unseeded defects into truth/fixtures #36: seeded bugs (caseN-b*) byte-identical in behavior; patches must apply cleanly and round-trip.

Success criteria

• A fresh engine run on the updated fixtures produces no finding that is neither truth-matched nor a deliberate bait, with no known-item exclusion list.

Open questions

• u2: keep as truth, demote, or split? (needs the transcript reasoning first)

Decision log

• Registered at C2 close per the deferral rule (spawned-by C2: argus methodology improvements (argus-review repo) #29).

[SSFSKIM]

[board] ready-for-agent: before the next scored run-id (C3/C4 benches)

[SSFSKIM]

[bench data 2026-07-27] case4-u2 is now a pure cross-model disagreement, and the fixture pass gains 8 new candidates.

• u2 (remote umask gap): missed by EVERY Claude engine — argus plain r3/r4, argus high r2 (actively refuted), builtin /review, builtin /code-review medium — while codex finds it every run. Both builtins caught only the LOCAL face of the same relocation (fresh archive dir born 0700). The refuter-transcript pull planned here should now also weigh whether u2's truth entry needs its remote/local faces split.
• New genuine-unseeded candidates from the builtin round + plain r4 (details in tests/review-bench/results/2026-07-27-c4-attribution-scores.json): case1 watermark read-only property; case2 per-promo ceil stacking over-credit; case2 resolvePromos param narrowing; case4 archive dir born 0700; case4 du-failure swallowed (need_kb=0, platform-divergent under pipefail); case4 replica-failure destroys local backup (design-call); case5 retained-exception traceback pinning vs the 512 bound; case5 _lookup mutates on probe-only paths.
• Scoreboard context: codex 20/20, /code-review medium 19/20, argus high r2 19/20, /review 18/20, argus plain v0.4.2 18/20 — FP 0 everywhere.