watch: accept math-proof:siege (detector false positive, #240)

[askalf]

Triage of #240 (math-proof:siege). Verdict: benign, detector false positive. Same clause-spanning class as #220 / askalf/redstamp#170.

What I read

• Anthropic first-party plugin, in-tree at ./plugins/math-proof, fetched at anthropics/claude-plugins-official@ab024cd. skills/siege/SKILL.md sha256 f40134d3…6b542, byte-identical to the watch triage: math-proof:siege — flagged, awaiting review #240 snapshot.
• Flagged line SKILL.md:612 is in the wave-planning brief: at most two queries may go to "routes the main line does not descend from (set aside, not refuted, under summary rule (ii))". "set aside, not refuted" is the skill's own term for an obstructed route that was not refuted, defined in summary rule (ii) at :515-518. "summary rule" is a cross-reference to that rule. Nothing is being overridden.
• Whole-plugin check: no override or system-prompt prose; the only URL host is apache.org (LICENSE); scripts/ledger.py imports only re/sys/pathlib.

Why it still fires
truecopy pins @askalf/redstamp v0.7.5, where the gap is [^.]{0,40}. redstamp v0.8.0 (#181, closing #170) changed it to [^.,;:|]{0,40}, and with that regex this line no longer matches (checked locally). So no new redstamp issue is needed. Moving truecopy's pin to v0.8.0 should keep this class from re-flagging.

Verification
npm ci + node support/watch-accept.mjs <skill-dir> --files attributes the single finding to SKILL.md and the remainder scans clean. The emitted hash matches the snapshot.

Closes nothing by hand: per the issue, recording the acceptance auto-closes #240 on the next watch run. Source ticket: WT-86cdee52.

[sprayberry-redline]

Verdict: approve. The acceptance is scoped to the recorded SKILL.md hash, and its rationale explains how the detector crosses a clause boundary. The existing remainder scan keeps changed bytes and findings in other files subject to detection.