Release Axon v0.2.1 · jyunming/Axon

What's Changed

Dead dirty flag — _incoming_rel_dirty was set but never read, so the O(N×M) edge-count scan ran unconditionally on every call. It now short-circuits immediately when the index is clean.
Missing config field — query_cache_ttl was consumed via getattr(..., 1800) silent fallback but never declared in AxonConfig. It is now a proper field (default 1800 s) settable in config.yaml.
Persist executor data race — _persist_executor fell through to the shared 8-worker self._executor when AxonBrain was active, allowing concurrent entity + relation graph saves to race on _gr_persist_hashes. It now always uses a private single-worker pool, serialising all persistence jobs.
Shallow-copy persist race — entity and relation graph snapshots were shallow-copied (dict(...)) before being handed to the background write thread. Nested dict/list mutations on the main thread could corrupt the on-disk graph. Both now use copy.deepcopy under _graph_lock.

O(1) entity matching via an inverted token index — replaces the O(n) full-graph Jaccard scan per query (10–100× speedup at 50 k+ entities).
BFS traversal cache — LRU cache (512 entries, 15-min TTL) keyed on (entities, max_hops, hop_decay); invalidated on graph mutation and project switch.
Async graph persistence — entity/relation saves offloaded to a single-worker background thread pool; _flush_pending_saves() for safe shutdown.
batch_search() on the vector store — replaces N separate search() round-trips for multi-query / HyDE variants with one batched call.
Pre-rerank filtering — candidate pool sliced to top_k × 3 before cross-encoder scoring; up to 66% fewer rerank calls.
Query cache TTL — cache entries now store (timestamp, response); stale entries expire after query_cache_ttl seconds (default 30 min).

README.md is now the single source of truth for both GitHub and PyPI descriptions (PYPI_README.md removed).
AxonCopilot.gif optimised from 20.5 MB → 3.25 MB (below PyPI/Fastly 5 MB proxy limit).
Live CI + PyPI version badges added to README; Quick Start now shows pip install axon-rag.
bump_version.py now auto-rebuilds VSIX + refreshes Cargo.lock in one shot.
audit_packaging.py checks 6 version-bearing artefacts (added bundled VSIX filename check).
Removed 29 obsolete one-off bench/debug scripts from scripts/.
pyproject.toml URLs updated (studio_brain_open → Axon); bundled VSIX updated to axon-copilot-0.2.1.vsix.

Full Changelog: v0.2.0...v0.2.1