Knowledge density scoring: surface high-signal sessions first

[ashu17706]

What

Score sessions by knowledge density and use it as a recall ranking signal. High-density sessions (decisions made, bugs fixed, code committed) should rank above low-density sessions (exploration, Q&A, back-and-forth) when recall scores are close.

Signal sources (all already in smriti tables)

Signal	Table	Weight
Tool calls with success	smriti_tool_usage	+1 per tool
File operations (writes)	smriti_file_operations	+2 per write
Git commits	smriti_git_operations	+3 per commit
Errors resolved	smriti_errors	+2 per resolved error
Session tags (decision/*, bug/fix)	smriti_session_tags	+5 per quality tag
Session cost (proxy for depth)	smriti_session_costs	log-scaled

Schema addition

ALTER TABLE smriti_session_meta ADD COLUMN density_score REAL DEFAULT 0;

Computed on ingest, recomputed by smriti enrich --density.

Recall integration

In recallMemories(), final score = relevance * 0.8 + density_normalized * 0.2. Configurable via SMRITI_DENSITY_WEIGHT env var (default 0.2, set 0 to disable).

Why

Two sessions with equal relevance scores are not equal. A session where you fixed a critical bug and committed a fix is worth more than a session where you explored an idea and abandoned it. The sidecar tables already capture this signal — we just don't use it in ranking.

No LLM calls needed. Purely structural. Ships before Phase 3.

Acceptance Criteria

• density_score computed on all new ingests
• smriti enrich --density backfills existing sessions
• Density factored into recallMemories() final ranking
• smriti show <id> --density explains the score breakdown
• Weight configurable via env var, 0 = disable

[ashu17706]

Implemented and committed on the dev branch.