Highlights
State-aware capsule retrieval — remembered capsules are now projected into the retrieval graph and a new state channel selects WHICH memory wins: items tagged state:<flag> partition above the rest while user_state.context_flags[flag] is active (lexical tie-break inside the group; state:calm + thematic coherence when no flag is active). Measured: 15/15 state-appropriate-recall on a 5-scenario × 3-state eval (factual mode; 12/15 auto) vs a 5/15 ceiling for state-blind systems. Validated on that set — external claims pending a holdout.
Keyword recall ranked by relevance, not recency — a fresh capsule matching one weak term no longer outranks an older capsule matching every term (set-recall@8 0.55 → 0.85).
Assertion layer sensors — temporal-accuracy / staleness / contradiction eval gate; mention_count corroboration; optional paraphrase claim matching (PULSE_PARAPHRASE_CLAIMS=1).
Also: access-frequency salience (Phase A+B, off by default), opt-in near-duplicate capsule consolidation (dry-run by default), inert procedural-memory scaffold (content-guarded), MCP assertion validation.
Migrations 028–032 (additive). Frozen v3 emotional scoring untouched throughout. Full changelog: CHANGELOG.md (#30–#43).