chore: post-rename polish — close the deferred items and a required-check gap

[sethbang]

Closes the loose ends the venice-py rename surfaced and deliberately deferred. No rename work here — the repo rename (Phase 5) is separate and still deferred.

__version__ had a fallback that made its own test inert

__init__.py fell back to a hardcoded "2.2.0" inside except Exception, and 2.2.0 was the version being shipped. So test_version_matches_installed_metadata — written specifically to catch a broken metadata lookup — passed identically whether the lookup worked or raised.

Fixed structurally rather than by bumping the literal:

• fallback is now 0.0.0+unknown, which can never equal an installed version
• handler narrowed to except PackageNotFoundError, so unrelated failures stop being swallowed
• new test_version_fallback_is_not_a_plausible_release greps the fallback out of the source and fails if it is ever shaped like a release

Verified by flipping the fallback back to 2.2.0 and confirming the new test fails, then restoring it — the gate is not vacuous.

The imports had to stay inside the try block (E402/I001), which is why the original was written that way.

Every Python directory in the repo is now linted

make format-check covered src/ and tests/ only.

• tools/ was already clean, so widening lint / format / format-check was free.
• benchmarks/ had 92 ruff findings, so it is staged across two commits — clear the findings first, extend the gate second, because doing it in one step would have turned the required validate job red.

All 92 were non-behavioural: 66 blank lines carrying whitespace, 6 missing trailing newlines, 5 unused imports, 5 unsorted import blocks, and the modern-syntax rewrites (List/Dict/Optional → builtin generics and X | None, a redundant open mode, one getattr with a constant attribute — whose hasattr guard and type: ignore are preserved).

No test covers benchmarks/, so verification was an ignore-whitespace diff review plus compiling and importing all 8 modules. examples/ was already covered by examples.yml.

configure snapshotted the config path at import

from ..config import DEFAULT_CONFIG_PATH bound the value at import, so redirecting the config location reached venice_ai.cli.config but not the configure command — the fixture had to patch both modules. The real hazard was the next test author patching only one and silently writing to the operator's live ~/.venice-py/. Now read at call time; duplicate fixture patch and its now-unused import removed.

The module alias is required — configure() binds a local config that a bare from .. import config would shadow.

Changelog

Four version headings (2.0.1, 2.0.2, 2.1.0, 2.2.0) were written as [x.y.z] links with no link definition, so they rendered as literal bracketed text. Added, along with an [Unreleased] section.

The 2.2.0 entry also described the bridge in the future tense — "a final venice-ai release follows this one… until it does, venice-ai stays at 2.1.0". Both clauses resolved when venice-ai 2.1.1 shipped, so the entry was stating something false.

One decision recorded in docs.yml

--project-name=venice-ai-docs names a live Cloudflare Pages project, not this package. It is not user-visible, and Pages projects cannot be renamed in place. Without a note saying so, a future rename sweep reads it as a stale venice-ai reference and breaks the docs deploy. The comment is the fix.

Verification

make format-check (now src/ tests/ tools/ benchmarks/), make check-all, mypy on 190 files, vulture, bandit — all clean. tests/cli/ + tests/unit/test_client.py: 1448 passed.

Full make test is the operator's to run.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[sethbang]

Update: this PR found a gap in the local gate itself

The branch went red on pyright (project) — one of the five required checks — while make check-all, make format-check, mypy, vulture and bandit were all green locally.

Root cause: pyright appeared in no make target at all. check-all chained format, lint, type-check (mypy), check-imports, check-dead-code and security. So a full local run could pass on code a required check rejects, and the first signal was a red PR.

mypy does not stand in for it. The finding was:

error: "PackageNotFoundError" is possibly unbound (reportPossiblyUnboundVariable)

The name was imported inside the try block that its own except clause names — so if that import had ever been the failing thing, the handler would raise NameError instead of handling anything. mypy does not report that class; pyright does. It was written that way because moving the import out at the bottom of the module tripped E402; the correct fix is to put it with the other module-scope imports, where E402 does not apply and the name is always bound, leaving only the version() call in the try.

Fixed both: the binding, and the missing type-check-pyright target now wired into check-all — verified by reintroducing the exact bug and watching the new target fail.

make check-all now covers all five required checks.