YantrikDB

TL;DR

Two real regressions I introduced in v0.7.13/v0.7.14. Both fixed and verified locally against the actual rejection surface before this release.

If you tried set_embedder_named() on v0.7.13 or v0.7.14 and got "after extract, expected files missing": v0.7.15 fixes that. If you tried pip install yantrikdb --no-binary :all: (build from sdist) on v0.7.12/v0.7.13/v0.7.14 and got "no matching distribution": v0.7.15 also fixes that.

The two bugs

#26 — set_embedder_named() never extracts (introduced in v0.7.13)

PR #23's refactor moved the tarball extraction loop into a new extract_tarball_to() helper, then deleted the inline loop. The second Edit's old_string accidentally included the call-site extract_tarball_to(&bytes, &tmp_dir)?; along with the dead comment block being deleted. Result: the helper was defined + unit-tested but never invoked from fetch_and_extract. Every set_embedder_named() download in v0.7.13 + v0.7.14 created the tmp_dir but never extracted anything — silently empty cache dir, then cache_is_populated() returned false and the error message said "expected files missing."

Worse than the original #15 bug (which only affected multilingual): v0.7.13/v0.7.14 broke ALL named-embedder downloads, including potion-base-8M and potion-base-32M that had been working since v0.7.2.

Fix in commit: restored the call + added a cleanup-on-error arm so a half-written tmp_dir doesn't survive a partial extract.

Sdist LICENSE path (v0.7.12 / v0.7.13 / v0.7.14 publish failure)

v0.7.12 switched the publish workflow from PyO3/maturin-action@v1 upload to pypa/gh-action-pypi-publish@release/v1, which strictly validates PEP 639 license metadata. PyPI started rejecting sdists with:

400 License-File LICENSE does not exist in distribution file
yantrikdb-X.Y.Z.tar.gz at yantrikdb-X.Y.Z/LICENSE

v0.7.14 attempted a fix by copying LICENSE into crates/yantrikdb-python/ (where maturin's manifest-path lives). But local maturin sdist + tar -tzf inspection — which I didn't run before claiming v0.7.14 fixed it — shows the copy landed at the nested path yantrikdb-0.7.14/crates/yantrikdb-python/LICENSE inside the sdist, not at the sdist root where PyPI's metadata validator expects it.

v0.7.15 fix: declare license-files = ["LICENSE"] in [project] of the root pyproject.toml. PEP 639-compliant approach tells maturin to package the listed file at the sdist root. Locally verified by running the exact PyPI-rejection-surface command:

$ tar -tzf yantrikdb-0.7.15.tar.gz | grep -i license
yantrikdb-0.7.15/LICENSE                          ← here ✓
yantrikdb-0.7.15/crates/yantrikdb-python/LICENSE  ← duplicate, harmless

What this means for users

Pre-v0.7.15:

| Operation | v0.7.13 / v0.7.14 |
|-----------|-------------------|
| pip install yantrikdb (wheel) | ✅ works |
| pip install yantrikdb --no-binary :all: (build from sdist) | ❌ "no matching distribution" |
| db.set_embedder_named("potion-base-8M") | ❌ "expected files missing" |
| db.set_embedder_named("potion-multilingual-128M") | ❌ "expected files missing" |
| db.with_default() (bundled embedder) | ✅ works |

v0.7.15: all four work.

Discipline reflection

Both regressions shared the same root cause: I claimed each fix shipped without running the exact command that catches the bug.

• For #26: unit tests for the helper passed but the integration path was never exercised.
• For LICENSE-in-sdist: cargo build + cargo test were green but maturin sdist + tar -tzf (the actual PyPI rejection surface) was never inspected.

The "user-side smoke" discipline I've been claiming to adopt applies here, but I narrowed it too much. The new procedural rule (saved to memory): when fixing a specific failure mode, the verification command must be the actual command that fails — don't substitute "near" verification (compile success, unit test pass) for "far" verification (actual rejection surface).

Verification (this release)

Full local CI sweep + actual rejection surfaces, all green BEFORE push:

• tar -tzf yantrikdb-0.7.15.tar.gz | grep LICENSE → file present at sdist root
• cargo fmt --all -- --check: PASS
• flake8 src/ tests/ --select=E9,F63,F7,F82: 0 issues
• pylint src/yantrikdb tests/ --disable=C,R,W --enable=E,F: exit 0
• cargo clippy --all-targets --all-features --workspace --exclude yantrikdb-python -- --cap-lints warn: clean warnings
• cargo test --workspace --exclude yantrikdb-python: 1422 default / 1411 slim
• pytest tests/: 220 passed
• maturin develop --release: clean install
• CI on PR #27: 3 required checks GREEN + windows-latest GREEN + macos-14 GREEN (confirms PR #25's macOS-14 failure was the rustup-init infra flake, not our code)