feat(dependabot): enforce AI-gated auto-merge for major github-actions bumps

[bbkrr]

Flips the pilot from observe to enforce: AIGATE_DRY_RUN "true" → "false" (one line).

After this, for realrate/RealRate-Private a major github-actions Dependabot bump whose AI verdict is auto gets approved + auto-merged — still gated by the repo's required CI (verify-output, pre-commit, semgrep, pip-audit, Analyze), so a bump that breaks CI never merges. human/uncertain verdicts still route to manual review.

Why this is safe to enforce now

• Verdict engine validated live against GitHub Models gpt-4.1 (the model fixed in the prior PR, since gpt-5 is unavailable for inference):
  • docker/metadata-action v4→v6 (real #2226) → auto / low ✅
  • breaking-change fixture (removed/required inputs, changed defaults) → human / high ✅
  • benign Node-runtime-only bump → auto / low ✅
• Defense in depth: github-actions ecosystem only · AI must say low-risk/non-breaking · required CI still hard-gates the merge · default-to-human on any uncertainty.
• Fail-closed: if the Actions token can't reach the model, the script emits verdict=human and the bump routes to manual review — never a bad merge.

Note on observation

A live CI dry-run couldn't be forced: @dependabot recreate fires a synchronize event, which the gate intentionally skips (it runs once, on opened). The engine was therefore validated directly against real GitHub Models rather than via a CI dry-run. The first real opened major github-actions bump will exercise the Actions-token path end-to-end; fail-closed makes that safe.

Revert = flip back to "true". Ref RealRate-Private#2229.

🤖 Generated with Claude Code