chore(release): bump version to 0.25.0

[scarletkc]

What this PR does

1. Version bump 0.25.0 (vexor/__init__.py, plugin manifest, server.json) — merging to main triggers the automated tag + PyPI publish + GitHub release.
2. Roadmap housekeeping: removed shipped items (hybrid implementation, .vexorignore, project-local cache, config/provider split); the P0 hybrid entry is now the pending default-flip gate with current benchmark status; project-level config split into a whitelisted v1 (existing Config fields only, credentials/endpoints rejected, directory-aware load_config) and a v2 (per-project scan defaults) that waits for v1 adoption.
3. Docs cleanup (findings from a full read of README/docs/skill/AGENTS after the four feature PRs):
   • README: removed the two feature-detail sentences from "Why Vexor?" (features belong in docs; README links instead), rerank example now shows hybrid neutrally, CLI Reference pointer mentions ignore files and --local.
   • docs/cli.md: the rerank paragraph sat under "Common Flags" but rerank is a config setting — replaced with a pointer to Configuration → Rerank.
   • docs/configuration.md: "This release raises the schema version…" was release-bound wording inside evergreen docs — replaced with a timeless cache-migration note under "Where data lives"; fixed a redundant phrase.
   • docs/mcp.md: the hybrid reranker/score note was stranded in the client-setup section — moved next to the vexor_search response schema.
   • docs/development.md: removed a reference to _label_for_path, which no longer exists; points at the vexor/modes.py strategies instead.
   • text.py messages reviewed — no changes needed.

Tests

python -m pytest tests -q (.venv, offline): 549 passed. Docs + version metadata only; no runtime changes.

---

Release notes for v0.25.0 (to attach to the GitHub release after the automated publish)

Highlights

• Hybrid retrieval (opt-in) — vexor config --rerank hybrid fuses full-corpus BM25 with dense ranking via weighted reciprocal rank fusion (0.7 dense / 0.3 BM25, k=60). BM25 term statistics are persisted in the index, so exact identifiers and lexical-only matches outside the dense candidate window become retrievable. Tokenization keeps whole identifiers (load_index_vectors) alongside sub-tokens. Scores under hybrid are normalized rank-fusion scores in [0, 1], not cosine similarities. A reproducible benchmark ships in scripts/eval_hybrid.py; the default ranking is unchanged (off).
• .vexorignore — per-project ignore rules with full gitignore semantics (per-directory files, ancestor chain, !negation). Always honored, including with --no-respect-gitignore; applied after .gitignore so it can re-include gitignored paths.
• Project-local indexes — opt in with vexor index --local, index(..., local=True) (Python), the new local argument on the vexor_index MCP tool, or simply mkdir .vexor; the project''s index (including its embedding/query caches) then lives in <project>/.vexor/index.db. Vexor writes a self-ignoring .gitignore inside so the database cannot be committed by accident. Explicit cache_dir/data_dir overrides take precedence; config, FlashRank assets, and local models stay global.
• Provider capability metadata moved to vexor/providers/capabilities.py; vexor.config re-exports everything, so imports are unchanged.

Upgrade notes

• Index caches rebuild once. The index schema version changed (v7, BM25 statistics); existing indexes are invalidated and rebuilt automatically on next use. Cached embeddings are reused, so the rebuild re-extracts but does not re-embed unchanged content.
• If you parse MCP/JSON scores: with reranker: "hybrid" the score field is a rank-fusion score in [0, 1], not a cosine similarity.

🤖 Generated with Claude Code