perf(security-issue-sync): load stage-specific rules only when they apply - #1997

Summary

• Moves the stage-specific Step 2b bullets of security-issue-sync out of signals-to-actions.md into siblings, behind pointers that name their triggers, and trims the Step 5b prose that push-decision already decides.
• Fixes two gaps found by comparing that prose with the code, so the cut does not change behaviour.
• Adds a pointer-routing eval step for the new pointers.

Moved byte for byte:

• release-manager-comments.md: the fill-fields, release-manager hand-off and publication-ready comments.
• post-advisory-records.md: the advisory-records update and the security-pages reminder.

Code fixes:

• Failed state fetch. After a push, a fetch that failed or returned unknown fell through to "fire the fill-fields comment". The skill says to post nothing on the hand-off front and retry on the next sync; push-decision now does that, with a recap blocker.
• Version noise in titles. Wrong label in the UI in Apache Foo 2.1.0 and Wrong label in the UI (Foo 2.1.0) passed the title gate, although the gate promises no version-noise suffix.

Type of change

• Skill change (.claude/skills/<name>/) — eval fixtures updated below
• Python package (tools/*/ with pyproject.toml)

Test plan

• prek run --all-files --skip workspace-pytest passes, including lychee
• workspace-pytest passes for the changed files
• 4 new test cases failed before the fixes and pass now
• Validator exits 0; no surface_hash changed
• security-issue-sync evals: 45/45 before, 53/53 after, no original case regressed
• New pointer-routing step: 8/8 on the real skill; with both pointers deleted, all 6 cases that need a file fail and the 2 that need none pass

The output spec names no file, so the agent has to get each name from the skill text.
case-5 covers an existing marker comment with no label transition, the easiest trigger to miss.

RFC-AI-0004 compliance

• Vendor neutrality — placeholders kept; the check-placeholders hook passes

No new mutation, host access, outbound message or LLM routing.

Linked issues

None.

Notes for reviewers

• security-cve-allocate changes too. The version-noise fix is in the shared strip(), so that skill now removes these suffixes as well. This matches its own rule (no redundant version suffix), but it is a behaviour change in a second skill.
• The Step 5b trim is the least tested part. Nothing checks that an agent acts on push-decision's reason the way it followed the deleted prose, and those strings are terser, for example the expired-session reminder.
• The verify step still only runs for allocated records. The removed prose said "when the pushed JSON carried review-ready", under the heading of the allocated → review-ready gate. The code's narrower reading is unchanged.
• Less than the earlier 8–10k estimate. The four judged gates and the push mechanics must stay, the comment bodies are already template files, and the Step 5d sweep needs adopter configuration and judgement, so no render or apply-plan commands were added.
• pointer-routing records a stated decision. The agent says which files it would open; it does not run with file tools.
• Evals predate the rebase. They ran before rebasing over generate-cve-json: warn when a credit line joins several names by comma #1741, which touches none of these files; prek and workspace-pytest ran after it.
• Single model. No cross-model pass and no adversarial review were run.
• tools/spec-loop/.last-sync is not bumped. The lifecycle spec is updated by hand here.

🤖 Generated with Claude Code