fix(nac): canonicalize tool signatures + substring-scan keyword query

[dennys246]

Summary

Fixes the predictions=0 in every enrichment_trace finding from the post-#218 sim audit. Two related bugs:

Phase 1 — Storage canonicalization. Three NAc storage formats coexisted for the same logical tool:

• tool_dispatch.py / tool_pain_bridge.py stored tool:<name> (canonical, via build_tool_signature)
• planning_bridge.py stored <name> (bare, drift)

Persisted aut_nac.json from a verification sim showed both forms in _links for every tool that ever took a plan-outcome path:

['tool:rusty_sword_slash', 'rusty_sword_slash', 'tool:say', 'say', ...]

Same logical tool, two dict keys, queries hit at most one. planning_bridge (observe + predict) and memory_agent._build_causal_context (query) all migrated to route through build_tool_signature, which is documented as "the single source of truth for tool→NAc event signature format."

Phase 1.5 — Substring scan in bio_enrichment._query_nac. NAc stores compound signatures like tool:rusty_sword_slash, but bio_enrichment extracts narrative keywords like rusty, sword, slash from percept text and called get_links_for_event(kw) — an exact dict lookup that never matched anything across the three pre-fix sims (predictions=0 in every enrichment_trace).

Added NAc.scan_links_for_keywords as a public companion to get_links_for_event: case-insensitive substring containment, dedupes by link id, sorts by confidence, drops short stop-words below min_keyword_length. bio_enrichment._query_nac now delegates to it.

Verification

A 5-turn embodied sim (qwen2.5-14b-instruct via leader) before vs after this branch:

	Pre-fix	This PR
_links keys	['tool:rusty_sword_slash', 'rusty_sword_slash', ...] (dual-form)	['tool:rusty_sword_slash', 'tool:sense_tools'] (canonical only)
enrichment_trace.predictions	0 in every event	1 in 2 of 4 events

Out of scope, deferred

• Phase 2/3 (substrate _links_by_node bridge) is paused. Substrate state inspection across 4 sims showed _reward_bias is fully empty in tool-only sims by design — distribute_reward only fires on Reaction events (pain / valence), which don't trigger from plain tool dispatches. Bridging _links to substrate node_ids would expose data that doesn't exist for typical sims. Revisit when a damage-taking combat sim verifies _reward_bias populates somewhere first.
• _query_atl (concepts=0) has the same exact-match shape as the bug fixed here; same fix pattern would apply but it's not blocking the predictions path.
• agent_pool.py "{agent_id}:respond" is a deliberately different format for NPC turn outcomes; no other writer or reader uses build_tool_signature shape for it. Documented in place.

Test plan

• Full fast suite: 6336 passed (was 6307; +29 new tests)
• ruff check + ruff format on every touched file
• New tests:
  • 4 scan_links_for_keywords cases (substring match / dedupe-and-sort / short-keyword drop / confidence floor)
  • 1 record_plan_outcome_uses_canonical_signature regression in test_memory_hub
  • 4 existing test_bio_enrichment tests updated to mock the new query method
• Verification sim showing predictions > 0 and clean _links shape

🤖 Generated with Claude Code