v0.4.0 - retrieval observability triangle

Highlights

The retrieval pipeline is now self-describing end-to-end: observe, dry-run, synthesize.

Features

• REMEMBER signal telemetry (#150). Every node returns per-signal scores (_remember_score, _vector_sim, _bm25_score, _recency_score, _graph_score, _co_bonus, _recall_boost, _rank_stage). Result.meta["signals"] carries fusion weights + per-stage candidate counts + reranker state + nucleus state. Always on, additive.
• SYS EXPLAIN REMEMBER (#151). Dry-runs the pipeline without materializing nodes, running rerank, expanding nucleus, or mutating recall counts. Returns candidate plan with per-signal scores + same telemetry. Safe for iterative tuning.
• ANSWER verb (#152). ANSWER "q" [USING "reader"] + q.answer(...). Retrieves via REMEMBER + hands context to a user-configured reader LLM. Returns {answer, cited_slots, candidates, reader}. Graphstore ships zero LLM deps; reader is a plain callable wired at construction. Named-reader registry for A/B.
• Skill-guided LLM ingest adapter (#149, #154). Bench harness for Mem0-style LLM-at-ingest experiments with cross-session fact memory + raw LLM dump per session.

Infrastructure

• Skills split (#148). graphstore-dsl (runtime DSL emission) + graphstore-builder (Python adapter authors) replace the old 9k-token monolith.
• Docusaurus docs (#142-147). Landing at https://graphstore-docs.orkait.com.
• LoCoMo bench CLI flags (#154). --adapter {graphstore,skill} and --use-raw-turns for proper A/B control.

Breaking changes

None.

Reader example

def my_reader(prompt: str, max_tokens: int = 1000) -> str:
    ...  # call any LLM backend

gs = GraphStore(reader=my_reader)
r = gs.execute('ANSWER "What is the capital of France?" LIMIT 3')
r.data["answer"]        # "Paris"
r.data["cited_slots"]   # ["mem:paris", "mem:eiffel", ...]
r.meta["signals"]       # full pipeline telemetry

Verified

• 1766 tests pass, 101 skipped
• Docusaurus builds clean
• Em dashes: zero (Rule 9)

What's NOT in this release

• Full LoCoMo A/B of baseline vs skill-adapter (abandoned; LoCoMo ships pre-distilled observations so re-distilling isn't a real test)
• Cloudflare Pages CI deploy workflow (manual-deploy only)

Both land in a follow-up once real user signal arrives.

🤖 Generated with Claude Code