v0.4.0

Added

• MUNIN_LLM_BASE_URL env var (default https://openrouter.ai/api/v1) (#123). Makes the LLM endpoint configurable for both the answer-quality eval and the consolidation worker, so they can target a local OpenAI-compatible server (llama.cpp / Ollama / vLLM) instead of OpenRouter. The API key becomes optional when a non-default base URL is configured — a local inference server that needs no auth can be targeted without supplying a dummy key. Default behavior (unset → OpenRouter, key required) is byte-for-byte unchanged. The two previously-separate OpenRouter clients (eval + consolidation) are now unified behind src/internal/openrouter.ts, preserving consolidation's exact request shape.
• temporal_stale advisory maintenance signal (#114). memory_orient and memory_attention now flag a tracked projects/*/clients/* active status when it references a YYYY-MM-DD date that has already passed while the surrounding phrasing is still forward-looking ("going to … <date>"). Advisory only — it surfaces the entry for the owner to restate and never rewrites content. memory_attention gains an include_temporal_stale filter (default true).
• Retrieved-but-never-followed-up quality signal (#99). memory_patterns (a new retrieved_unused pattern) and memory_orient (maintenance_needed) now surface entries that are repeatedly shown in search results but never opened or acted on — a signal that they may be stale, mis-tagged, or low-value. Owner-only, scoped to projects/*/clients/*, over a rolling 30-day window with a ≥5-impression / zero-follow-through threshold. getInsightsByEntry gains an optional since window and restrictToTracked flag. An observe-only precursor to learned reranking.
• GET /heimdall.json service self-descriptor endpoint (#135) for Tier-1 discovery by the Heimdall dashboard.

Fixed

• memory_read_batch now records retrieval outcomes (#99). It read entries without logging an opened_result outcome (unlike memory_read), so entries retrieved via search and then bulk-read showed zero follow-through in retrieval analytics. Batch reads are now counted as follow-through, fixing a latent gap in memory_insights and the new retrieved-unused signal.