Ruff selection is documented org-wide but enforced per-repo, inconsistently

[gavinbee]

Follow-up to #36, which fixed this repo only. #36's root cause — a documented lint selection that no config file enforced — is not unique to .github. Verified against each repo's default branch on GitHub (not local checkouts):

Repo	pyproject.toml	select	Lints in CI?
rems-sync	yes	["E","F","I","W"]	ci.yml
deck-eval-gen	yes	["E","F","I","W"]	ci.yml
.github	no → added by #36	—	ci.yml
deck-eval-parser	no	—	no — tests.yml runs pytest only, and ruff isn't in requirements.txt or requirements-dev.txt
officials-admin	no	—	no workflows at all (CI arrives with #12)
swim-club-tech-survey	no	—	no — only run_survey.yml

Two distinct problems fall out.

1. Enforcement gap — lints, but with no config

A repo that runs ruff check . without a config uses the installed ruff's defaults, and with ruff>=0.6 those can change under it. That is #36 exactly: main green, next unrelated PR red. .github was the only repo in this state, which is the embarrassing part — the repo that publishes the standard was the one not following it.

2. Coverage gap — doesn't lint at all

deck-eval-parser and swim-club-tech-survey are outside the documented quality gate entirely. CONTRIBUTING.md § Quality gates says "Python repos use ruff", and for these it simply isn't true. deck-eval-parser has real code and its own tests.yml that never calls the shared reusable workflow.

Worth deciding whether that's intended (survey repo may be a scheduled script, not a maintained codebase) or drift. officials-admin is a known-pending case, not drift — #12 adds CI.

Why this can't just be "share the config"

Ruff has no remote config inheritance. extend takes a local path, so there is no way for six repos to point at one file in this one. Options, all with tradeoffs:

1. Copy the config into each repo, plus a conformance check. reconcile-repo-defaults.yml already heals settings drift on a schedule; it could equally assert that every Python repo has a config whose selection matches the documented one. Keeps per-repo autonomy, catches drift, but the config itself is still duplicated.
2. Centralize the flag in reusable-python-ci.yml — add a ruff-select input defaulting to the documented selection, so every consumer is deterministic without touching each repo. Recommended for enforcement. Caveat: a CLI --select overrides file config wholesale, including per-file-ignores, so a repo needing extra rules must pass the input rather than configure locally. No conflict today, since the two repos with config already match the default.
3. Ship the config in a pip package and extend into site-packages. Genuinely shared, and fiddly enough that it probably isn't worth it at six repos.

1 and 2 compose well: 2 makes CI deterministic now, 1 keeps local ruff check . honest for developers, who don't run the workflow.

Acceptance criteria

• Every repo that contains Python either lints with the documented selection or has an explicit, recorded reason not to.
• Local ruff check . and CI agree — a developer can't be green locally and red in CI, or vice versa.
• A new Python repo created via scripts/create-repo.sh starts out conformant, rather than relying on someone remembering to copy a config.
• CONTRIBUTING.md § Quality gates describes what is actually enforced.

Known instance to fix in passing

officials-admin's requirements-dev.txt (added in swimblocks/officials-admin#16) declares no ruff, because that repo has no lint gate yet. Whatever this issue lands should cover it, or #12 should.

[gavinbee]

Decision: 88 with E501 enforced, and run ruff format

Recording @gavinbee's decision so #38 has a target rather than an open question:

• line-length = 88, everywhere.
• E501 enforced — the ignore = ["E501"] in deck-eval-gen and rems-sync comes out.
• ruff format runs as part of the gate, not just ruff check.

What the repos look like today

Surfaced while working #46, where two marginal lines failed lint in this repo that would have passed in four others.

Repo	line-length	select	ignore	E501 fires?
.github	unset → 88	E,F,I,W	none	yes, at 88
deck-eval-gen	100	E,F,I,W	E501	no
rems-sync	100	E,F,I,W	E501	no
deck-eval-parser	no config	ruff defaults	—	no
swim-club-tech-survey	no config	ruff defaults	—	no

Three behaviours, none of them chosen on purpose:

1. .github is strict for two compounding reasons — it selects all of E without the E501 ignore the other configs pair it with, and it never sets line-length, so it inherits ruff's default 88. The repo publishing the standard ended up 12 columns tighter than the repos following it.
2. deck-eval-gen and rems-sync set 100 and switch E501 off.
3. The two with no config escape for a third reason: ruff's default selection is E4/E7/E9/F, and E501 is E5, so it is never checked.

Verified against ruff 0.16.1: a 139-character line passes ruff check --isolated and fails under --select E with Line too long (139 > 88).

line-length = 100 is currently inert

Worth calling out because it looks like a decision and isn't. E501 is the only rule that consumes line-length, and both repos that set it also ignore E501. CI runs ruff check ., not ruff format. So in deck-eval-gen and rems-sync the 100 has no effect on anything — those repos have no effective line limit. Whatever we land on, a line-length sitting next to an E501 ignore should not survive: it is a claim the tooling does not back.

Why 88

It is ruff's default and Black's, so ruff format assumes it with no configuration, and any contributor arriving with default tooling matches. PEP 8's own limit is 79, with an explicit escape hatch to 99 for teams that agree; 100 and 120 are common but have no standard behind them. 88 also makes this repo's current behaviour correct on purpose instead of correct by accident.

The counter-argument, and why it does not win

The existing configs justify the ignore as "many existing strings are long on purpose" — URLs, prompts, error messages that read worse wrapped. That is real, but it argues for # noqa: E501 on the specific offenders, not for switching the rule off across a repo. ruff format also will not split string literals or comments, so long strings stay long and stay visible as deliberate exceptions rather than hiding inside a blanket ignore.

Scope this adds to #38

Beyond the select inconsistency this issue already covers:

• line-length = 88 in every Python repo's config
• Drop ignore = ["E501"] from deck-eval-gen and rems-sync
• # noqa: E501 where a long line is genuinely intentional
• ruff format in the gate — decide --check in CI vs formatting on commit
• The two repos with no config get one; the two with no lint job get one (already noted above as the coverage gap)

Consequences per repo — how much ruff format rewrites, how many E501s appear once the ignore is dropped — are deliberately not measured here; that belongs to the work itself.