build(tooling): run the hooks with prek and manage Python deps with uv

[Minipada]

Mirrors the move already made in ~/dev/monorepo: .pre-commit-config.yaml is now run by prek — same file, same schema, one Rust binary that resolves hook environments with uv and needs no Python environment of its own — and dependencies live in a PEP 621 pyproject.toml + uv.lock instead of poetry's [tool.poetry] + poetry.lock.

No issue for this one; it came out of the observation that the formatting gate had stopped running at all.

The formatting job was dead, not slow

.github/workflows/format.yaml installed poetry and then pip-installed requirements.txt + requirements-dev.txt just to get pre-commit on PATH. That install had been failing outright — ERROR: Failed to build 'fastcov' … ModuleNotFoundError: No module named 'pkg_resources' — so its last two runs never reached a single hook, and the workflow was then disabled manually. Every finding below is therefore pre-existing drift the dead gate was hiding, not new. (The workflow is enabled again as of this PR, which is where it is first exercised: 24 hooks, 25 seconds.)

The job now installs prek and nothing else: no poetry, no pip, no apt. clang-format arrives in the mirror hook's own wheel, check-xml uses Python's parser rather than xmllint, and graphviz was only ever there for pydeps, which nothing calls.

Python lint/format is ruff only

flake8 + 13 plugins, black, isort, pycln, pyupgrade, flynt, yesqa and the python-no-log-warn pygrep hook are replaced by ruff-check + ruff-format — one binary instead of ~20 hook environments pre-commit had to build before linting a line. [tool.ruff.lint]'s selection is chosen for parity with what the old stack enforced (E/W/F, I, UP, B, C4, A, Q, N, G/LOG, T10, EXE, and RUF100 for yesqa), not for maximum strictness: ruff's newer opinions (the rest of RUF, SIM, PTH, …) are deliberately left off, the same "broaden over time" stance the monorepo's first rollout took. setup.cfg's [flake8]/[darglint] sections are gone; darglint has no ruff equivalent worth turning on blind and is not replaced.

Fallout, all mechanical, fixed rather than ignored: 16 findings from the parity rule set (unsorted imports, a stale # noqa: T201 ×3 for a rule this repo never enabled, datetime.utcnow, three (str, Enum) classes → StrEnum, two bare zip()s → strict=True, three root-logger calls in the dashboard → a module logger, one dict comprehension) and 14 files reformatted by ruff-format (mostly black 23's missing blank line after a module docstring). Two more the dead job had hidden: dc_triggers' two .cpp files were never clang-formatted, and tools/e2e/Containerfile's find … | xargs tripped hadolint DL4006/SC2038 — rewritten as find … -exec … +, which drops the pipe instead of silencing the rule.

Dependencies

requires-python moves from the humble-era >=3.8.10,<3.9.7 || >3.9.7 to >=3.12 (Jazzy/Ubuntu 24.04), with .python-version pinning 3.12 so a local uv sync builds against the interpreter the code actually runs on. Deliberately held below 2: numpy (Jazzy's rclpy/cv_bridge are built against 1.26) and pydantic (draw_image.py and the dashboard still use v1 APIs — BaseSettings).

Dropped from the dev group as used by nothing: ipython, pydeps, pylint, cmake-format, plus strictdoc and fastcov, which live in the containers that actually run them (containers/doc/Containerfile, tools/e2e/Containerfile) — fastcov's sdist is what broke CI in the first place.

requirements.txt and requirements-dev.txt are deleted, not re-exported: pyproject.toml + uv.lock are the single dependency source, and a generated copy in the tree is only somewhere for the two to disagree. doc/src/dc/setup.md and the two AWS demo pages now say uv sync (--no-dev for the runtime set alone), and .dockerignore allows pyproject.toml + uv.lock in their place.

The ROS manifest hooks are vendored

minipada/ros2_pre-commit's ros2_ws_same_versions finds setup.py with an unbounded find "$SRC" -name setup.py, so once uv sync puts a virtualenv at ./.venv it checks every setup.py in site-packages — numpy's, for one — against the DC version. Both its hooks also shell out to xmllint, which GitHub's runner image doesn't ship, and that was the only system dependency left in the whole hook set.

tools/ci/pre-commit/ros2_package_checks.py does both checks over dc_*/ only (versions: one version across every package.xml and each dc_*/setup.py declaring it; metadata: no package left the template's placeholder license, maintainer or description), reading manifests with ElementTree — the same parser the check-xml hook already uses. stdlib rather than defusedxml: the input is this repo's own manifests, and the dependency would mean giving the hook a Python environment it otherwise doesn't need.

Also: a comments rule in CLAUDE.md

File headers one to three lines, block comments one or two, why not what, long rationale in progress.txt / an ADR / the PR description. Prompted by this branch's first draft of format.yaml, which carried a 22-line header essay. ci.yaml, doc.yaml and tools/e2e/Containerfile predate the rule — to be trimmed when next touched, not copied.

Verification

• CI: Formatting (prek) passes on this PR — all 24 hooks in 25s, hadolint's docker-image hook included.
• prek run --all-files --skip build-doc passes locally too (adds build-doc, which doc.yaml owns in CI).
• uv sync --frozen resolves and installs on 3.12; every dev-group and runtime package imports, and ruff from that venv is the same 0.16.3 the hook pins.
• All 63 tracked .py files byte-compile. check_if_numbers_are_consecutive and the dashboard's three StrEnums were smoke-tested directly (.value, == against a plain str, and iteration order all unchanged).
• The vendored checks were negative-tested against a scratch workspace at the same depth: a mismatched package.xml version, a mismatched setup.py version and a package with all three placeholder fields are each reported, and both checks pass once the workspace agrees (single-quoted version='…' included, which upstream's grep did not accept).

🤖 Generated with Claude Code

https://claude.ai/code/session_015aZ8Ng8bK4mK9keCKxPzwX

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 67.54%. Comparing base (cfb94dd) to head (78a3fbc).
⚠️ Report is 1 commits behind head on jazzy.

Additional details and impacted files

@@            Coverage Diff             @@
##            jazzy     #350      +/-   ##
==========================================
- Coverage   67.55%   67.54%   -0.00%     
==========================================
  Files          95       95              
  Lines        5885     5884       -1     
==========================================
- Hits         3975     3974       -1     
  Misses       1910     1910              

Flag	Coverage Δ
cpp-jazzy	67.54% <100.00%> (-<0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.