feat(review): require CodeGraph for structural evidence in parallel subreviews

ci-review-prompt.md

@@ -71,7 +71,11 @@ run concurrently), one per evaluation dimension group:
    naming and reserved-word safety, repository conventions, performance.
 5. experience — UX surfaces, DX surfaces, visual/DOM, accessibility/i18n.
 Give each dispatch the changed files and surfaces it must inspect and require
-source-backed path:line findings. Treat subagent output as evidence, not
+source-backed path:line findings. Require every dispatched subagent to use the
+configured CodeGraph MCP tools for its structural questions — callers/callees,
+impact radius, dependency and test reachability, base-vs-head flow — before it
+concludes, and to cite the CodeGraph query it relied on; grep-only structural
+claims are not sufficient when CodeGraph is reachable. Treat subagent output as evidence, not
 authority: independently verify any blocker you adopt, resolve conflicts
 against source, and write the final control block yourself — every approval
 gate in this contract still applies to the synthesized result. Skip a

code-reviewer-prompt.md

@@ -7,6 +7,12 @@ reformat code, create commits, push branches, or change configuration. You may
 suggest exact code changes or minimal patch snippets only when they clarify the
 fix; the primary agent or developer must make any change.
 
+Use the configured CodeGraph MCP tools aggressively for structural evidence:
+call graph and callers/callees of changed symbols, impact radius, dependency
+and test reachability, and base-vs-head flow comparison. Prefer CodeGraph over
+grep for any structural claim and cite the query you relied on; fall back to
+direct file inspection only when CodeGraph is unreachable, and say so.
+
 ## Prime directive
 
 Review the changed code with high signal. Find issues that materially affect