The workflow ran commit-check-action, which installs a released
commit-check -- so every pull request was checked by the version before
it. #540 had its own title rejected by the bug it was fixing, and the
same thing happened tonight in the .github repository. A self-test that
cannot see the change under test is not a self-test, and the action's
version does not track the engine's, so a fix needed two releases and a
bump before it could be reproduced here.

Install from the checkout instead. The cost is job-summary and
pr-comments, which are the action's rendering rather than the engine's
behaviour and are covered by the action's own repository.

Enumerating the commits is not optional. On a pull_request checkout HEAD
is the synthetic merge commit, which the engine skips, so a bare
'commit-check --message' reports a pass having read nothing. Verified on
a scratch repository holding one plainly bad commit: the bare form exits
0, this workflow exits 1. HEAD^1..HEAD^2 gives the real commits, and a
guard fails loudly if HEAD is not a merge so the failure can never be
silence.

The pull request title reaches the CLI through the environment. Titles
are attacker-controlled, and interpolating one into a run: block is a
script injection.

SonarCloud's unpinned-install rule failed the quality gate on a bare
'pip install .'. main.yml and publish-package.yml both carry
--only-binary :all: for exactly this, added in #479; this step was the
one that did not. Verified it still installs a local path: the resulting
version reads 2.13.4.post1.dev4, which is the checkout rather than the
release.

With stdin left open, commit-check waits to read a message even when
only --branch and --author-name were asked for. Reproduced locally: the
process runs until killed, which in CI is a step that hangs rather than
fails.

Closing stdin also decides what the author checks look at, so the
comment says which: the author of HEAD, matching what the action
reports today.

I added --only-binary :all: because main.yml and publish-package.yml
carry it and #479 mentions SonarCloud. It does not satisfy this rule --
the alert stayed -- and --upgrade pip added a second flagged line, so
the change took the count from one to two.

Back to one plain line. The remaining alert needs a decision rather than
another flag: main.yml has the identical install and passes only because
it is not new code, so locking would mean a hash-pinned requirements
file or reviving the stale uv.lock.

The session was already there and already wrong: it ran a bare
'commit-check --message', which on a pull_request checkout inspects the
merge commit, which the engine skips. Leaving it that way while putting
a correct copy in YAML would have left two implementations, with the
broken one being the one a contributor reaches for locally.

So the logic lives in noxfile.py and the workflow is one line. The same
command now reproduces a CI failure on a laptop, which was half the
reason for moving off the action.

The session adapts rather than assuming CI: HEAD^1..HEAD^2 when the
checkout is a merge commit, HEAD otherwise, and the title only when
PR_TITLE is set. It still refuses to pass silently -- missing HEAD^2
during a pull_request event is an error, not a fallback.

Verified locally: enumerates 2 of 2 commits on a merge ref and 0 on a
plain one, exits 1 on a bad title having still run the branch and author
checks, exits 0 on a good one, and no longer hangs on stdin.

Reverts b95acfd, which moved the pull request checks into the nox
session. The session is a developer command -- run commit-check on your
working copy -- and folding CI's shape into it made the simple thing
complicated for no gain.

The justification was wrong too. I claimed the same command would
reproduce a CI failure locally, then had to branch on whether HEAD^2
exists, because locally there is no merge commit and no PR title. A
command that behaves differently in the two places does not reproduce
one from the other.

It also introduced a bug the shell never had: filtering on .strip()
dropped empty messages, which this repository rejects via
allow_empty_commits = false, so such a commit would have been reported
as 'HEAD is not a merge commit' instead of as the thing it is. The shell
keeps them: printf 'a\0\0b\0' through 'read -r -d' yields a, empty, b.

noxfile.py is now byte-identical to main.

main is linear and every subject ends in (#N): this repository squashes,
so the commits on a branch never reach it. The title becomes the subject.
Checking each commit was protecting history that does not exist, and cost
thirty lines of shell to do it.

What is left is two commands. The bare 'commit-check --message' still
cannot be one of them -- it reads HEAD, the synthetic merge commit, which
the engine skips and would pass having read nothing -- so the title goes
in through stdin instead, which sidesteps HEAD entirely and needs no
enumeration.

Contributors lose CI feedback on intermediate commit messages. The
pre-commit hook is where that belongs anyway: it arrives while the
message is being written rather than a round trip later.

Verified on a merge ref: bad title 1, good title 0, good title with a
bad branch name 1.