feat(learn): add continuous knowledge consolidation layer

[ashu17706]

Summary

Smriti currently only captures sessions and can manually distill one batch of them via smriti share --segmented. Nothing persists locally unless a user runs share, and nothing tracks whether a piece of knowledge is a one-off or something that keeps coming back.

This PR adds a Continuous Knowledge Consolidation Layer, applying two ideas from Building a Second Brain (Tiago Forte):

• Progressive Summarization — cheap Stage-1 extraction (segmentSession) runs broadly over dense sessions; expensive Stage-2 polish (generateDocument) only runs once a unit proves it's reused (recalled repeatedly) or scored high relevance at extraction time.

• Intermediate Packets — the existing KnowledgeUnit type now has somewhere to live before export, in a new smriti_knowledge_units table, so reuse can accumulate.

• src/db.ts — new smriti_knowledge_units table + CRUD helpers (insertKnowledgeUnit, findUnsegmentedDenseSessions, findPromotableUnits, incrementRetrievalCount, promoteKnowledgeUnit, listKnowledgeUnits)

• src/search/recall.ts — tracks retrieval_count on every recall() exit path (best-effort, never breaks recall)

• src/learn/consolidate.ts — new consolidateKnowledge(): segment phase (dense, unsegmented sessions → Stage 1) + promote phase (units that cleared the reuse/relevance bar → Stage 2 → written to .smriti/knowledge/ + recorded in smriti_shares)

• src/index.ts, src/format.ts — CLI wiring for smriti consolidate and smriti learnings

• src/team/share.ts — exported getSessionMessages for reuse by the new module

CLI-only, not wired into the daemon — consolidation runs two LLM-driving stages, more than enrich, which is already excluded from the daemon's flush path for the LlamaCpp-crash reason documented in src/daemon/index.ts.

Test plan

• New test/learn-consolidate.test.ts covers: segment-phase content-hash dedup, promotion threshold (retrieval/relevance bar), graceful degradation (a per-unit write failure doesn't abort the run), and retrieval tracking via recall()
• bun test --cwd ./test — 378 pass, 0 fail across all 31 files, no regressions
• Manual smoke test: smriti consolidate --min-density 0 and smriti learnings run cleanly against a fresh DB

---

Generated by Claude Code