Initiative: MCP-powered CI review enrichment (optional, graceful-degrading) #659

Integrate optional MCP (Model Context Protocol) servers into the Claude Code CI review so reviews can draw on real-time documentation and external signal (deprecated APIs, breaking changes, secret/dependency scanning) instead of relying solely on training data. Opt-in only: when no servers are configured, behavior is byte-identical to today (zero impact on existing adopters). Degradation follows Fail-Loud-Never-Fake -- an unavailable/invalid server emits a ::warning:: annotation and the review proceeds, never a silent failure or a fabricated all-clear.

Grounding note (important): the source idea (discussion #650) names the central org reusable petry-projects/.github/.github/workflows/claude-code-reusable.yml. This repo deliberately does NOT use that reusable -- its review pipeline lives here: pr-review.yml (reusable-only) -> scripts/review-one-pr.sh -> scripts/engine.sh, where claude --print --model X --permission-mode acceptEdits --allowed-tools "Bash,Read,Grep,Glob" is the single CLI invocation. This plan therefore grounds the integration in engine.sh + pr-review.yml. See open_questions for the cross-repo scope decision.

Scope: (1) engine plumbing to pass an MCP config to the Claude CLI and expose its tools; (2) an optional mcp_servers input on pr-review.yml; (3) a regression + graceful-degradation guard; (4) per-stack server recommendation docs; (5) a human-gated pilot on one downstream repo, measured before any broader rollout. Source: #650

Planned from idea discussion #650 by the BMAD Scrum Master initiative-planner. Inert until a maintainer adds initiative:auto.