fix(ci): cache skill security scan results to avoid double LLM scans

[danbarr]

Summary

• When Renovate opens a PR bumping a skill, skill-version-check's autofix job pushes a follow-up commit that only touches spec.version. That retriggers build-skills.yml and re-runs the LLM-based security scan on source identical to what was already scanned, doubling OpenRouter cost per affected PR.
• Caches the raw scan output (skill-scan-<skill>.json, scanner-version.txt) in skill-security-scan, keyed on repository/ref/skill_path (what the scanner actually reads) plus a hash of scripts/skill-scan/requirements.txt. Check out skill source and Run skill security scan are skipped on a cache hit.
• The allowlist evaluation (process_scan_results.py, driven by security.allowed_issues in spec.yaml) is split into its own step that always runs, cache hit or not, against the possibly-cached raw scan output. This matters because adding/editing an allowlist entry is a common follow-up commit that doesn't change repository/ref/path — it must still get a fresh blocking decision, not a stale cached one.
• build-skill-artifacts's pre-flight scan gate still reads scan-summary.json from the uploaded artifact either way, so publish is still blocked on unallowlisted findings.

Test plan

• Confirm skill-security-scan restores from cache (and skips the LLM scan step) on the second pull_request synchronize run after a spec.version-only autofix commit — verified via this PR's smoke test: first run was a cache miss (full ~9min scan), re-running the same job hit the cache and skipped Check out skill source / Run skill security scan entirely, finishing in 31s
• Confirm a PR that only adds an allowed_issues entry still gets a fresh, correct pass/block decision even when the raw scan is served from cache
• Confirm a cache miss (new repository/ref/path) still runs the scanner normally — verified on the first run of this PR's smoke test
• Confirm build-skill-artifacts's pre-flight scan gate still blocks publish on a blocking finding — gate passed on both runs, reading the fresh scan-summary.json each time

🤖 Generated with Claude Code

[toolhive-release-app]

🛡️ Skill Security Scan Results

✅ agent-observability-eval-bootstrap

• Status: Passed
• Findings: 5
• Allowed (not blocking): 2
  • LLM_PROMPT_INJECTION (Allowed: FP: same as policy_violation above — all pattern matches are on documentation, code examples, or attack pattern descriptions for detection purposes, not agent instructions.)
  • MANIFEST_MISSING_LICENSE (Allowed: datadog-labs/agent-skills is licensed MIT at the repository root; upstream does not embed an SPDX license identifier in per-skill SKILL.md frontmatter.)

---

Summary: Scanned 1 skill(s), all passed security checks. ✅