[matthewp]

Changes

• Adds a new assessment-only capability, src/merge-risk/, that posts one scored risk comment on a pull request after Factory's own Factory Review check completes. It never approves, blocks, or merges anything.
• Scores five categories (alignment, impact, verification, rollback cost, error detection) from a bundled, non-configurable rubric (skills/merge-risk/SKILL.md, adapted from Warp's merge-risk skill) using a cheap, tool-free model call over PR metadata, the changed-file list, check conclusions, and review states — no repository checkout.
• The overall low/medium/high score is computed deterministically in trusted code from the model's good/ok/poor category verdicts and fixed weights, not trusted to the model's own arithmetic.
• Triggered only by check_run.completed on Factory's own review check (not on every push), so it only runs for pull requests that were actually labeled for review or had astro-reviewer requested — keeping cost down given PR volume.
• One comment per pull request: found and updated by a marker on every re-assessment rather than accumulating new comments.
• Wires in MergeRiskCoordinator/MergeRiskWorkflow (new Durable Object + Workflow binding, v10 migration) and a new router.ts rule that validates the check run's app id before enqueueing.

Testing

• No new automated tests; existing suite (459 tests) still passes with no regressions, since this is additive and the router/dispatch changes are pure branches gated on a new event/check name.

Docs

• Updated README.md with a "Merge risk" capability section, the architecture diagram, and the routing-rules summary. No external docs needed — this repo isn't published as a package.