SQLite snapshot storage, docs site, and release workflow

[atharvas]

Summary

• Storage backend rewrite: snapshots now live in a single SQLite database (<snapshot_dir>/snapshots.db) instead of thousands of per-test .pkl/.pkl.gz files. Payloads are pickled, gzipped, and content-addressed by sha256 in a refcounted blobs table, so benchmarks producing identical outputs share one blob on disk. Per-test metadata lives in a snapshots table.
• JSON sidecars unchanged: the per-test <module>/<class.benchmark>/<param-hash>.json metadata sidecars and the flat baseline.json are still written exactly where they were, since downstream tooling consumes them. store_snapshot now returns the sidecar path.
• Clean break: there is no .pkl/.pkl.gz read path. Existing pickle snapshot directories are not migrated — callers re-capture. Version bumped 0.1.0 -> 0.2.0 to signal the format break.
• Docs: new MkDocs Material site under docs/ (+ mkdocs.yml) and a CLAUDE.md describing the codebase.
• Release infra: .github/workflows/release.yml publishes to PyPI via OIDC trusted publishing on GitHub Release (no secrets).

Storage shape (shapely roundtrip, local smoke)

58 snapshots -> 45 unique blobs   (~22% deduplicated)
raw payload  111 MB  ->  gzipped in SQLite  84 MB

Test plan

• pytest unit suite (excluding tests/test_repos/, tests/test_cli_roundtrip.py, and the two pre-existing-hang discovery tests): 153 passed, 8 skipped
• tests/test_storage_comprehensive.py: 32 passed (incl. new dedup + DB-compression tests)
• tests/test_integration.py: 22 passed (fixed store_failed_capture to coerce non-string failure_reason)
• Local end-to-end shapely roundtrip: list -> capture -> baseline -> verify → 57 pass / 1 skip, correct transition matrix
• CI: tests.yml (py3.8–3.13) + test-astropy/pandas/shapely roundtrips

Notes

• The Lint and Format job is pre-existing red on main (failing on every recent commit, mostly ruff rules vs. the py3.8 Optional[...] convention and vendored tests/test_repos/ sources). Not introduced here; my changed files are ruff format-clean. The meaningful gates are the test workflows.
• tests/test_class_capture.py / test_class_instance.py hang on main independently of this change — they BenchmarkDiscovery into tests/test_repos/ (~1479 benchmarks). Out of scope here.

---

Update: CI repair folded in (per maintainer request)

CI was 100% pre-existing red on main (Lint failed in ~9s; Tests ran 2–6h then failed via an unbounded discovery hang). This PR now also makes CI genuinely green:

• ruff config (pyproject.toml): exclude vendored tests/test_repos/ (third-party sources), ignore the pyupgrade rules that conflict with the deliberate py3.8 Optional[...] convention, and per-file-ignores for the intentional __init__.py import-after-logging pattern and test-scaffolding rules.
• real lint fixes: removed dead locals + tightened a strict type check in comparator.py, set-comprehension in transitions.py, repo-wide ruff format.
• discovery hang: test_class_capture.py::test_class_instance_capture walked all of tests/ (incl. ~1500 vendored benchmarks) and ran each with no timeout. Scoped it to its test_class_instance fixture module + added a 30s runner timeout.

Local verification of the exact CI commands:

• ruff format --check src/ tests/ + ruff check src/ tests/ → All checks passed
• pytest -v --ignore=tests/test_repos/ --ignore=tests/test_cli_roundtrip.py → 155 passed, 8 skipped in ~2.6s (was failing after hours)