Fix scoring engine: check bucketing, path classification, blocker language, GitLab merge status

[JackSpiece]

Fixes correctness bugs in the triage engine that were producing wrong advice. No public API, dataclass shape, or CSV/render contract was changed, so render.py, cli.py, and the CSV columns are untouched.

What was wrong

1. summarize_checks silently dropped legacy commit statuses

GitHub's statusCheckRollup mixes CheckRun (status + conclusion) with StatusContext (state only). StatusContext entries matched no branch, so a repository whose CI reports through commit statuses (CircleCI, Jenkins, Buildkite, most non-Actions CI) was scored "no visible checks" while it was bright red. That is the single worst bug here: the tool told maintainers a broken PR was safe to review.

2. Cancelled runs were scored as failures

CANCELLED and STALE carry no verdict. Reporting them as failures made the tool tell maintainers to chase authors over runs that were superseded by a newer push or manually stopped.

3. Unknown conclusions bucketed as pending

Any conclusion the code did not recognise fell into pending, so choose_action returned "wait for CI" forever for runs that will never report again. Unknown/verdict-less conclusions now bucket to skipped.

4. ACTION_REQUIRED was treated as a failure

It means a human still has to approve or resume the run. It is pending, not broken.

5. An all-skipped or all-cancelled run read as "CI passed"

The no visible checks branch keyed off checks.total, which counted skipped runs. It now keys off passed + failed + pending.

6. File classification matched bare substrings

This is the bug with the widest blast radius, because code_files and test_files feed the code changed without tests (+10) and no test plan found (+8) flags.

Path	Was classified	Now
src/latest_release.py	test (contains test)	code
src/contest_manager.py	test (contains test)	code
src/license_checker.py	doc (contains license)	code
src/readme_generator.py	doc (contains readme)	code
src/generated_ids.py	generated	code
src/build/compiler.py	generated	code
deploy/main.tf	uncounted	code
Dockerfile, Makefile	uncounted	code
.github/workflows/*.yml	uncounted	code

Classification now anchors on whole path segments, basenames, stems, and extensions, and the four categories are mutually exclusive and evaluated generated > test > doc > code, so a file can no longer be counted as both a test and code. Build-output directory names (build, out, target, coverage, .next) only count at the repository root.

User-supplied test_hints / doc_hints / generated_hints from config deliberately keep their historical substring behaviour, so no existing .maintainer-radar.json changes meaning.

7. Bot comments were scored as maintainer feedback

Codecov posts "3 checks failed". Dependabot posts "this PR is blocked". Both matched BLOCKER_RE, added +25 risk, and forced the PR to "needs author follow-up" with the message "respond to unresolved maintainer feedback" — for a comment no maintainer wrote. Bots are now detected by login suffix ([bot], -bot, _bot), by a known-bot list, and by type / __typename / is_bot.

8. Pasted CI logs were scored as maintainer feedback

A maintainer quoting a failing log, or pasting it in a code fence, tripped the same +25 flag on the words in the log. Code fences, HTML comments, quoted (>) lines, and inline code are now stripped before matching.

9. GitLab ci_must_pass and not_approved mapped to CHANGES_REQUESTED

Neither means a human asked for changes. ci_must_pass is a pipeline gate; not_approved means review has not happened yet. Both were adding +25 risk and routing the merge request to "needs author follow-up" when the author owed nothing. Only discussions_not_resolved maps to CHANGES_REQUESTED now. Both conditions are still visible through mergeStateStatus: BLOCKED (+6) and the pipeline check.

10. GitLab manual-gate pipelines

blocked, canceling, and waiting_for_callback were missing from the pending set, so a manual gate fell through to an unrecognised conclusion instead of reporting as pending.

Deliberate non-changes

• has_body = "body" in pr is unchanged, on purpose. An earlier read of this called it a bug — that was wrong. A shallow scan omits the body key entirely, while a real PR with an empty description has body: "". Treating an empty string as "no body" would waive the missing-test-plan penalty for authors who wrote nothing at all. The current check is correct.
• .json is not in CODE_EXTENSIONS. Most JSON in a diff is fixtures, lockfiles, or generated schema. Counting it as code would fire code changed without tests on data-only PRs. Specific generated JSON names are already handled by GENERATED_BASENAMES.
• CheckSummary and FileSummary field names are unchanged, because render.py and the CSV exporter consume them through __dict__.

Tests

24 new tests, 48 total. New classes: CheckRollupTests, FileClassificationTests, BlockerLanguageTests in tests/test_scoring.py, and GitlabMergeStatusTests in tests/test_normalize.py. Every bug above has a test that fails on main.

Also wraps a 116-character line in analyze_pr that exceeded the configured line-length = 110.

Still outstanding (not in this PR)

• _render_markdown_row in render.py interpolates raw title_text into a Markdown table cell while the adjacent _markdown_pr_label correctly routes through markdown_cell(). A PR title containing | or a Markdown link breaks or injects into the generated report. Left alone because render.py and its 32 KB test file were out of scope for this change.
• mypy is still continue-on-error: true in CI, so type regressions cannot fail a build.
• ci.yml has no fail-fast: false, so one matrix leg failing hides the others.
• No v0.21.0 tag exists despite __version__ = "0.21.0" and the README pinning @v0.21.0.
• maintainer-radar is not published on PyPI, so the README's pip install maintainer-radar does not work.

[JackSpiece]

Closing this because I am no longer pursuing this work. The branch and commit history are preserved.