[anth-volk]

Fixes #77

What this does

A follow-up to #75, which merged at 0deef32 before its second review's fixes. 16.1 · 17 in the plan. The noqa-comments guard now counts the suppression forms it still missed, and the rule, README and registry guide say what the check does: the total number of inline suppressions may not grow. The baseline stays at 140 entries in 49 files, because the repository has none of the newly counted forms.

• Directives beside a trailing file-level noqa (bug 1, medium). Ruff reads each kind of directive in a comment on its own. A trailing ruff: noqa or flake8: noqa that Ruff does not apply after code used to end the parse, so print(1) # noqa: T201 # ruff: noqa: E501 had no entry. Now the other directives in such a comment count, and so does the file-level one when it is on its own line.

• A leading comma (bug 2, low). Ruff skips a comma before the first code, with a warning, so # noqa: ,T201 counts as T201.

• The isort comments (bug 3, low; review round 1 of this pull request). These switch off Ruff's import sorting (I001), so they are now entries, labelled by form, read exactly as Ruff 0.16.10's extract_isort_directives reads them:

  • Whole comments: skip_file in its four spellings (# isort: skip_file, # isort:skip_file and the same after # ruff: ), and # isort: off or # ruff: isort: off, with trailing space removed.
  • Skip markers: any other comment containing isort: skip or isort:skip, which counts only when an import statement starts on the comment's line. That covers # keep first, isort: skip_file, # isort:skip_file # keep first, # isort: skipped, ... and # isort:skip_this.
  • Not counted: an own-line skip, a skip after code that isn't an import, #isort: off and # isort:off. Ruff applies none of them. on isn't counted either, and inside an off range Ruff reads no other isort comment.

• The rule id stays emitter.quality.inline-suppressions-only-shrink. Renaming it now would make the rules tool treat it as a new rule and restart its comparison, the loophole the second review found. Compare a renamed ratcheted rule's baseline with the base branch #79 tracks making the tool follow a renamed ratcheted rule.

• Wording. The rule statement, README, baseline header, the guard's module docstring and its messages now say:

  • one entry per directive;
  • the total may not grow against the base branch, so a pull request may add a suppression only by removing another;
  • the check counts directives only, so it doesn't stop a swap at the same total, a rename or a code added to an existing comment.

  The README's list of forms includes flake8: noqa and the isort directives.

• The registry guide now matches its own list of what only a person may do:

  • an agent may remove suppressions and shrink the baseline;
  • an agent adds a baseline entry only as part of a swap at the same total, such as a new suppression in place of one it removes, and that entry needs a person's approval in review. Nothing enforces that approval today: main requires no approving review and there is no code-owners file. Otherwise, an agent that can't fix a finding without a new suppression stops and reports it;
  • raising the total is a person's decision, because the guard refuses a grown count with or without a [rule.downgrade];
  • renaming a ratcheted rule's id restarts its comparison, so review catches an id change (Compare a renamed ratcheted rule's baseline with the base branch #79).

• Tests. The three same-total tests now model a pull request: the base branch's files and baseline, then the pull request's. They check the guard on both and the total against the base branch's, and they assert the baselines differ. The higher-total test is built the same way and fails as before. The base-branch comparison receives only the two totals, because check_baselines compares counts, so these tests don't show what a comparison of entries would do.

• One baseline reader. file_length, ci_scripts and pyright now read their baselines with baseline.baseline_lines, as UTF-8, like the guard and validation, so the helper's docstring holds.

Separate issues: #78 for the two timing-dependent macOS tests, and #79 for the rule-rename limit.

Left for later, from this pull request's review:

• The baseline writers (--write-baseline and the like) don't name an encoding.
• # isort: split isn't counted, since it divides sorting without switching any of it off, and the statement excludes it.

The [rule.downgrade] section naming #74 and the 140-entry baseline, both from #75, still await Anthony's confirmation; this pull request doesn't change either.

Test-first record

#75's commits, now on main, include 29f5ad6 (red) and a5bdc0c (green), which replaced b9bd74f and 9c76f80 when #75's branch was rebased.

Commit	Result
2396b3a Tests for the three bugs	Red, on assertions: all 11 fail. Five are directives beside a file-level noqa: the guard returned no entry, or only the file-level one. One is the leading comma, which returned no entry. Five are the isort forms, which returned no entry.
a71ecd8 The fixes	Green: all nine test classes pass by name, including this repository's baseline at 140. One code comment that named the isort skip keyword was itself read as a directive, and is reworded.
5585b92 One baseline reader	The rules guards pass.
65825b5 Same-total tests modelling pull requests	They pass at the same total; the higher-total case fails with "baseline grew from 1 to 2".
fba91c2 The wording and the guide	render updated the rules views; all six guards pass with GITHUB_BASE_REF=main, and guard --base-ref origin/main passes. ruff check passes on the changed Python files.

Review round 1 of this pull request:

Commit	Result
0ef3d36 Tests of isort as Ruff reads it	Red, on assertions: all 8 fail. The four skip samples got no entry, and the four forms Ruff ignores got one.
e13f908 isort read as Ruff 0.16.10 does	Green: all ten test classes pass by name, including this repository's baseline at 140.
f01aa5e Test of the "directive" message	Red, on its assertion: the leftover-entry message still said "such comment".
d2c392b The guard's docstring and messages, and the guide	Green: all ten classes pass. The six guards pass with GITHUB_BASE_REF=main, and ruff check passes on the changed files.

The work is in a fresh clone; the old checkout had lost objects.

Only the test file's classes were run, each by name with python -m unittest; no test suite or directory was run.

CI

No workflow, trigger or script changes, and no environment variable or secret is read differently.

Release

The fragment is changelog.d/noqa-ratchet-follow-up.changed.md, a patch bump: no shipped code changes, so the wheels are unchanged.

🤖 Generated with Claude Code