v0.4.0

[0.4.0] - 2026-05-22

Added

• Preference lane v2 — preference-shaped drawers now store a
  preference_span metadata field and optional pref_embedding secondary
  vector. Preference queries receive a separate preference_match score so the
  known single-session-preference weakness can improve without adding an LLM
  to the default path.
• Query intent provenance — searches classify queries as preference,
  decision, how_to, definition, temporal, or unknown, and return the
  intent in MCP and Palace::search_with_provenance results.
• Optional local rerank path — PALACE_RERANK=1, palace search --rerank,
  or palace_search({"rerank": true}) reranks top hybrid candidates and returns
  rerank_score provenance. The default search path is unchanged.
• Gain v2 feedback — palace_gain now accepts an optional record payload
  for explicit usefulness feedback and reports precision_at_1,
  precision_at_5, and per-intent precision. Diary entries that cite returned
  drawer IDs also infer useful feedback automatically.
• Eval CI lane — CI now runs the coding-agent memory eval separately and
  checks the sampled LongMemEval baseline guard.

Changed

• palace upgrade-embeddings --refresh-preferences and
  palace_upgrade_embeddings({"refresh_preferences": true}) can refresh
  preference-span embeddings while re-embedding drawers.
• Palace::search_with_provenance is additive: results include
  preference_match, intent, and rerank_score.