feat: insights table — session-agnostic synthesized knowledge #69

Concept

Add an insights table that stores synthesized knowledge discovered when multiple references are brought together — distinct from both raw summaries and explicit user preferences/decisions.

Key characteristics

• No session_id — insights are corpus-wide, not scoped to a conversation
• Derived, not captured — an insight is produced by reasoning across multiple sources (documents, summaries, older turn patterns), not by directly summarizing one turn
• Different from decisions/preferences — a decision is "always do Y"; an insight is "these three things relate in an unexpected way" or "this pattern appears across multiple sessions"
• Primary sources: document lookups (manuals, references), cross-session pattern recognition, recurring themes in work

Proposed schema

CREATE TABLE insights (
  insight_id  INTEGER PRIMARY KEY,
  text        TEXT NOT NULL,
  embedding   BLOB,
  tags        TEXT NOT NULL DEFAULT '',
  source_refs TEXT,          -- JSON array of (table, id) pairs that produced it
  model_id    INTEGER,       -- model that synthesized it
  timestamp   TEXT NOT NULL
);

source_refs links back to the documents/summaries that were combined — provenance for later review or re-synthesis.

Generation triggers (TBD)

• Explicit agent call (agent notices a connection and stores it)
• Post-document-import pass (new doc may relate to existing summaries)
• Periodic synthesis pass over recent turn summaries across sessions

Recipe integration

insights becomes a recipe layer kind, surfaced in general_search results and as its own optional layer. Since insights have no session scope, they're always candidates for the session-agnostic sweep.

Relationship to general_search (#65)

general_search queries summaries + documents. Once this table exists, it should also query insights. The distinction: summaries are compressed history; insights are synthesized understanding.