test(ci): make the slow-tier partition a partition, and check it (#672)

[willhea]

Addresses the partition half of #672, via option 3: keep the parallelism and the matrix, remove only the hand-maintenance. The matrix half is deliberately untouched.

The problem

The slow tier is divided between CI jobs by listing every module by name in ci.yml, and that partition is restated a second time as CI_SLOW_MODULES in tests/conftest.py. Neither copy was derived from anything, and neither was checked.

It had already drifted. tests/test_manifest_report_fixtures.py was listed in two jobs, so it ran once per leg of each: eight times per CI run.

The existing guard, test_every_slow_module_is_run_by_a_workflow, is a good one and is one-directional by design. It catches a module that runs nowhere. It is satisfied by a single naming and cannot count them, which is why nothing in the repository reported the duplicate.

Which listing was the accident

Resolved from history rather than guessed, because the module's name fits more than one job and #672 notes that nothing states the grouping rule.

Commit	When	What it did
bfa5d606	Aug 4, 16:51	Added it to the general slow list (now remaining-slow), deliberately, as part of fixing a gate that ran nowhere. The commit argues the placement at length.
f1767df9	Aug 4, 19:13	Added it again to corpus-gates, inside a commit titled "Fix conflict-resolution regressions from merging develop into #520" whose body never mentions CI.

So the corpus-gates entry is collateral from a merge resolution. That one line is removed. The module's 91 tests still run, once, via remaining-slow, confirmed by running them.

I had initially reasoned from the module's name that corpus-gates was the better home. History says otherwise, which is why it was worth checking.

The two guards

• test_no_module_is_run_by_two_ci_invocations — counts invocations per module. _modules_run_by_workflows returns a set, which is what makes it structurally unable to see a duplicate.
• test_every_ci_slow_module_is_named_by_a_slow_step — checks the second copy's own stated membership rule ("named by a slow CI step"), previously a claim about ci.yml that nothing verified. It matters because the tuple is read by pytest_runtest_logreport in every session, so a stale entry keeps applying a skip ceiling locally for a gate CI stopped running.

What this deliberately does not do

The issue also floats deriving CI_SLOW_MODULES from ci.yml. I'd advise against it and haven't done it.

That tuple drives the #288 skip ceiling, not the CI partition. Seventeen of the forty modules CI runs are absent from it, so deriving it would switch skip ceilings on for seventeen modules at once. That is a change to what the suite enforces, not a de-duplication, and it is precisely the decision #610 exists to make. The issue itself frames this one as "about the duplication that makes the gap easy to create, not the gap itself."

So the claim is asserted in the safe direction only: everything the tuple names must really be named by a slow step. The converse stays open and stays #610's to settle. The guard's docstring says so, so a later reader does not tighten it to an equality and make that decision silently.

Verification

Per the issue's instruction to confirm a new guard reddens on the tree as it stands, since the tree already contained the defect.

The duplicate guard is a real red-green, not a planted one. On unmodified develop it fails, naming the module and both jobs:

test_manifest_report_fixtures.py: ['ci.yml:corpus-gates', 'ci.yml:remaining-slow']

Removing the accidental entry turns it green.

The CI_SLOW_MODULES guard was mutated twice, because the first mutation did not prove it earns its place.

Mutation	Result
Delete a module from a slow step	Guard reddens — but so does the pre-existing coverage gate. Overlap, so inconclusive.
Change a step's marker -m slow → -m "not slow", module still named	Guard reddens alone: 1 failed, 22 passed. The older gate stays green because the module is still named by a workflow.

The two guards overlap partially and neither is redundant.

_slow_marker_selected is pinned directly, on the four marker spellings these workflows use, because the trap it exists for cannot be reached through the live workflows: the fast step is the one spelled not slow and not browser, and it names no modules, so a helper reading the bare word would classify it as a slow step with nothing going red. The wrongness would surface the first time someone added a module argument to that step. Simplifying the helper to "slow" in expression fails exactly that case and no other.

Gates. Fast suite 2015 passed / 4 skipped / 15 xfailed; ruff check and ruff format --check clean.

Caveat: measured in a worktree lacking the gitignored fetched corpus, which reports 4 skips where the main checkout reports 1. None relate to this change.

Notes for the reviewer

• CONTRIBUTING.md's "three steps", which Most checks on a pull request restate a verdict already given, and the hand-maintained split that produces them has drifted #672 also flags, is corrected on a separate docs branch (docs: correct three onboarding claims about running the tests #705) rather than here.
• This touches tests/test_ci_workflow.py, as does fix(ci): install and audit the committed dependency graph (#678) #700. Both edit the module docstring's property count, so whichever merges second will need a trivial conflict resolution there — the two additions are independent and both belong.
• The matrix half of Most checks on a pull request restate a verdict already given, and the hand-maintained split that produces them has drifted #672 (24 of 31 checks, four legs that have never disagreed) is untouched and still open.