Promote 2 deepseek-verified m3 findings into the golden set

[bborbe]

What

golden-curated-1 → golden-curated-2 (155 → 157 entries): two findings that MiniMax-M3 surfaced and deepseek-v4-flash-max[1m] independently verified as real defects are promoted to accepted.

entry	evidence (deepseek, runnable)
recurring-task-creator#30 — CI genuinely red	ran govulncheck: GO-2026-6179/6180 fire against golang.org/x/mod@v0.37.0, neither ID in VULNCHECK_IGNORE (Makefile.precommit:52), so make precommit / CI fails
discord-assistant#5 — apostrophe strip bug	reproduced exactly: "hey bot's weather" → "'s weather" via \b boundary + lstrip(" ,.:;-—") missing the apostrophe

Why this is a first for the set

Every prior entry was adjudicated from one model's (Opus's) opinion. These two are backed by two models' independent agreement — m3 found them, deepseek confirmed with runnable evidence — the exact evidence class the golden-set design specifies for promotion. This is the set's first step away from "recall measures agreement with Opus" toward "recall measures truth."

The aliasing check did its job

Four findings were deepseek-verified. Two of them (the unbounded uncaughtException swallow, the supervise-orphan) turned out to be re-statements of entries the set already held — the aliasing guard caught that promoting them would double-count, so they were dropped. Promotion would have looked correct and been wrong; the guard is why it isn't.

Score deltas (vs golden-curated-1 → golden-curated-2)

model	recall before	recall after	hits
opus (self-match)	0.844	0.832	114
m3	0.052	0.066	9 (+2)
deepseek	0.096	0.109	15 (+2)

m3 and deepseek each gained 2 hits (their own verified findings now count); opus's self-match dipped because the new entries are findings it missed — the tautology loosening, which is the point.

[ben-s-pull-request-reviewer]

The diff contains only CHANGELOG.md and bench/golden.json — documentation and benchmark data. No source code of any language is touched. No ast-grep YAML rules apply to these file types, and the Step 4b-i candidate query would produce an empty set for these non-code files. The mechanical funnel already ran with zero findings.

Per Step 4 early exit: no rule-relevant source files changed.

---

## PR Review: Promote 2 deepseek-verified m3 findings into the golden set

Verdict: approve

Summary: This PR adds two m3+deepseek-verified benchmark entries to bench/golden.json (vulnerability findings from recurring-task-creator#30 and discord-assistant#5) and updates the version stamp from golden-curated-1 → golden-curated-2. The CHANGELOG entry correctly describes the changes, the provenance chain, and the scoring impact. No code, no rule violations.

---

Must Fix (Critical): None.

Should Fix (Important): None.

Nice to Have (Optional): None.

---

Traceability (selector mode):

• Candidates: 0 judgment rules matched by glob filter (non-code files CHANGELOG.md, bench/golden.json match no source-language triggers)
• Applicable: 0
• Skipped: N/A — diff contains only benchmark data and changelog text; no source code rules fire

---

Step 3b (precommit): Skipped (selector mode) — CI covers lint+test.

---

Notes:

• The two promoted entries cite runnable evidence and include aliasing-check verification that they are not duplicates of existing entries. The provenance comment provenance: m3 finding, deepseek-verified 2026-08-18 (runnable evidence); genuinely new — NOT a re-statement of an existing entry (aliasing-checked) appears on both new entries and is consistent with the golden-set design intent.
• The discord-assistant#5 entry's line_when_seen: 864 in shim/claude_openai_shim.py refers to an external repo file; the finding is about the coding plugin's own golden.json benchmark, not about code in this repo.

---

{
"verdict": "approve",
"summary": "PR promotes two m3+deepseek-verified benchmark entries into the golden set with correct provenance, version bump, and changelog. No rule-relevant source code changed; all mechanical checks passed.",
"comments": [],
"concerns_addressed": [
"tests/correctness: two new golden entries cite reproducible evidence and aliasing checks confirming they are genuinely new findings",
"concerns: none — the two concerns from the plan (tests, correctness) are addressed by the entries' provenance documentation"
]
}