Harden 0.4.0: fix the [bracket] DDL crash, make the CLI scriptable, add CI

[dridk]

Why

python -m sqlite2duckdb chinook.db out.db crashed with Parser Error: syntax error at or near "[". Chasing that down surfaced a handful of other problems: the CLI could not run unattended, the Python floor was never applied, and there was no CI to catch any of it.

The crash

duckdb replays sqlite's index DDL verbatim through its own parser, which rejects the [bracket] quoting that chinook and MS Access exports use:

CREATE INDEX [IFK_AlbumArtistId] ON "albums" ([ArtistId])

The attached catalog is parsed on bind, so both the SCHEMA and the DATA halves of COPY FROM DATABASE fail, and the sqlite extension exposes no setting to turn index parsing off. duckdb_indexes() on the attached database raises the same error.

So COPY FROM DATABASE is gone. Each table is now recreated from the DDL duckdb derives for the attached catalog (duckdb_tables().sql), which always parses and still carries the primary keys and NOT NULL constraints that a CREATE TABLE AS SELECT would lose. The indexes are then read back from sqlite_master and their quoting translated.

This is a regression fix: the pre-0bc556a loop never touched indexes, so chinook worked in 0.3.x.

Also in here

Bugs. requres-python was a typo, so the >=3.8 floor never reached the wheel metadata. importlib.metadata.version() raised PackageNotFoundError on an uninstalled checkout. A failed conversion left a half-written .duckdb behind, which then tripped the "already exists" guard on the next run. Path arguments crashed. db_name came from an unfiltered duckdb_databases row and was interpolated without escaping.

API. sqlite_to_duckdb(src, dst, *, overwrite=False) returns a ConversionResult, raises FileNotFoundError / FileExistsError instead of bare Exception, accepts PathLike, and reports through logging rather than printing to stdout.

CLI. --force, --quiet, --verbose. The overwrite prompt now only appears on a tty; it used to hang or raise EOFError under CI, in a pipe, or under uvx ... < /dev/null.

Packaging. Everything moves to uv: dev dependency group (faker was imported by the tests but declared nowhere, so a fresh clone could not run them), committed uv.lock, requirements.txt dropped. The sdist gets an allow list, taking it from 592K back to 13K.

CI. ruff and pytest on 3.9 through 3.13, plus publish-on-tag through PyPI trusted publishing.

Issues

Closes #3 — that traceback died in SHOW tables on a northwind export, the same [bracket] cause. There is now a regression test for it.

Refs #1 — the reporter's exact information_schema.table_constraints query returns 12 rows on a converted chinook instead of none. UNIQUE, FOREIGN KEY and CHECK are still not copied, so this may or may not be enough to close it.

Refs #2 — COPY FROM DATABASE was adopted in 0bc556a and is removed again here, for the reason above. The simplification still stands: the USE / SHOW tables dance the issue pointed at is long gone.

Verification

8 tests to 32, including value fidelity rather than row counts alone, the error paths, the CLI end to end, and five integration tests against the real chinook now shipped in examples/. They compare against the sqlite source rather than hardcoding: same table set, same row counts, same values, every named index present.

CI is green on all five Python versions.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XtZtVXPQqXvdHLQ98bT2gF