Summary

Task Reference: [MYMR-207]

The three context-bundle builders (buildAgentContext, buildPlanningContext, buildReviewContext) built their Prerequisites and Downstream sections from raw recursive CTEs (getDependencyChain, getDownstreamTx) that predated the cancelled-transparent semantics buildEffectiveDepGraph already enforces for the analyze tools. As a result bundles listed cancelled tasks as prerequisites/dependents, and a cancelled middle on a depends_on path burned a depth slot so the real active blocking wall was never reached inside the maxDepth=2 budget (A -> B(cancelled) -> C(active) showed B at depth 1 and never reached C).

This change derives prerequisites and downstream from a depth-bounded walk over the effective graph so the four analyze tools and the three context bundles agree semantically for the same task:

• Extracted the traversal substrate (listTasksForGraph + listDependsOnEdges + adjacency build) into a new exported buildDepAdjacency in lib/graph/effective-deps.ts, and refactored buildEffectiveDepGraph to call it (behaviour unchanged; getBlockedTasks/getCriticalPath/deriveTaskStatesSlim see no difference).
• Added a new exported walkEffectiveDepsBounded: a BFS-by-effective-depth walk where cancelled tasks are transparent and depth-free, capped at the bundle budget of depth 2.
• Switched agent.ts/planning.ts/review.ts off getDependencyChain/getDownstreamTx onto one buildDepAdjacency call plus forward (prerequisites) and reverse (downstream) bounded walks. lib/data/traversal.ts is untouched; its other callers (getDownstream, edge.ts cycle detection) are unaffected.

Type of change

• Bug fix
• New feature
• Refactor / cleanup
• Documentation

Testing

• Tested locally with bun run dev
• Linting passes (bun run lint)
• Typecheck passes (bun run typecheck)

Full bun test suite green (461 pass, 0 fail). Added a cancelled-middle transparency integration test to each of tests/context/agent.test.ts, tests/context/planning.test.ts, tests/context/review.test.ts asserting C surfaces and B (cancelled) does not.

Notes for reviewer

• The bounded walk is intentionally capped at depth 2 (the existing bundle budget) while the analyze tools use the unbounded effective closure. That depth bound is the locked design decision: consistency here is semantic (cancelled-transparency, active-only nodes), not identical traversal depth.
• Tracked follow-up, deliberately out of scope: mymir_analyze type='downstream' (getDownstream -> getDownstreamTx -> fetchDownstream) still uses the raw cancelled-blind CTE and carries the same bug class. Left for a separate task per MYMR-207 scope; no change to getDownstream/getDownstreamTx/fetchDownstream/tool-handlers.ts here.