Deep split of the security skill family #1672

Part of #1342: the deep split of the security family, the same pattern as pr-management (#1570–#1585) and release-management (#1604–#1625).

The first pass (#1344) was a wording pass. The family is still the largest: 15 skills, 98,884 body tokens in total, 11 over the body budget and 7 over the 500-line cap.

Body tokens are the measured_tokens: stamps, which include the frontmatter and the generated pre-flight block (~850 tokens together), so the 5,000-token body budget corresponds to about 5,850.

Approach

• A new tools/security-management package holds the deterministic rules: classification, field extraction, tracker-body and comment rendering, and the scrub checks. The skills call it and keep only judgement and the confirmation gates.
• SKILL.md becomes a router; rules that serve one sub-action move into one sibling file per backend or outcome.
• Renderers are print-only. Nothing in the package posts, sends, labels or commits: every state change stays a proposal the user confirms.
• External content stays data: anything the package reads from a report, thread or PR is passed through as quoted values, never as instructions.

Stack (one native stacked PR per layer)

1. Package + the five import skills (issue-import, -from-pr, -from-md, -from-scan, -via-forwarder)
2. issue-triage + issue-invalidate
3. cve-allocate
4. issue-sync + issue-deduplicate
5. issue-fix

The model-* skills and tracker-stats-dashboard are under budget and stay out of scope.

Invariants

Same as #1342: headings and golden-rule headlines byte-identical, eval-extracted sections tightened only, every suite run before and after (outside the sandbox), skill-token-count, skill-surface-hash, the validator, prek run --all-files and lychee green.