Commit f8f4f38f31

tests: the review loop, driven end to end by three accounts

Approvals, CODEOWNERS, review threads and notifications each had a test.
Nothing put them together, so the loop a repository actually runs — draft,
ready, thread, approval, resolve, merge — had never executed once.

An author, an owner named in CODEOWNERS, and a third collaborator who
owns nothing. The third is what isolates the CODEOWNERS gate: their
approval satisfies require-approvals and leaves the owner requirement
unmet, so the refusal can only come from CODEOWNERS. All three settings
are on at once, which no single-feature test did.

This is not the second human #139 asks for, and the difference matters:
the CODEOWNERS bug was a gate that never fired, and a test written from
the same understanding as the code would have asserted the same wrong
thing. What it closes is the narrower gap that the features had never met.

Writing it found two things.

`mr ready` notified a thread's participants, which at that moment is the
author alone — the actor, excluded — so it reached nobody. It notifies
the repository's targets now, the way opening one does. That corrects the
choice made in the commit below this one.

And there is no way to ask a particular person for a review at all: the
queue is computed from involvement, and a collaborator with write access
who has not yet touched the thread gets nothing until they comment. Filed
as #145. The test asserts the current behaviour and fails if it changes,
so the issue and the code cannot drift apart silently.

Closes #139

Layout: unified · split