ci: gate what the corpus is worth, and retire a waiver's completed instruction

[fas89]

Two things that had quietly stopped being true.

1. mutation_coverage.py ran nowhere

run.py answers "do the cases pass" — a corpus of zero cases answers that perfectly. mutation_coverage.py answers the harder question: delete a constraint from the schema and see whether any case notices. It has been in this repo, runnable, documented, and wired into nothing, so the figure could have fallen to zero between releases in silence.

The floor is the current number rounded down (40.7% → 40), not a target. That is where mutation-testing gates converge: a break threshold set where you actually are catches a regression on day one, whereas an aspirational one goes red immediately and gets deleted by day three — worse than no gate.

Deliberately not a ratchet, and the divergence is on purpose. Ratchets exist to absorb run-to-run noise; there is none here, because the run is deterministic — mutate the schema, replay the corpus, count. A fixed floor does the same work with less machinery, and raising it stays a deliberate edit made when a corpus wave earns it.

Proven to fail, three ways:

Condition	--min-coverage 40
today's tree	exit 0
floor raised to 41	exit 1
metadata group trimmed to a single case	exit 1
group restored	exit 0

The third is the one that matters — it is a real corpus regression, not a moved goalpost.

2. A waiver told the next reader to fix something already fixed

compat-waivers.txt said docs/releases/0.7.2.md "should be corrected to name this change". It was corrected, in #46 — the note now opens "One breaking change, corrected here after the fact". The instruction outlived the work.

The waiver line itself stays, permanently: 0.7.2 is published and its history cannot be rewritten, so the gate has to be told this break is known rather than new. The comment now says that, and says what a waiver must never become — a way to quiet the gate about a break nobody dealt with. This one was dealt with by correcting the release note, not by adding the line.

Borrow-before-build

Searched:
- mutation testing CI gate ratchet vs fixed threshold → "set break thresholds, not
  aspirational targets"; "a gate that reports a baseline on day one and is disabled
  by day three is worse than no gate"
- JSON-Schema-Test-Suite conformance coverage / mutation prior art → the official suite
  organises by keyword category (this corpus already mirrors that, 11 groups); Bowtie is
  the cross-implementation reporter (check_reference.py is the local equivalent).
  No project publishes a mutation-coverage GATE for a schema corpus.

Reuse strategy: adapt-the-pattern (break-threshold-at-current), diverge on the ratchet.

Corpus 480/480 and meta-test 19/19 unchanged.

🤖 Generated with Claude Code

[fas89]

Once this lands I can set main to require all three contexts — gates can fail, corpus, backward compatibility — which is the change this PR exists to make safe.

Order matters and is not reversible cheaply: the contexts have to be produced by main before they are named as required, or every open PR hangs waiting for a report that will never come. So: merge this, confirm the three report on the next PR, then flip protection.

Two things that need your call at that point:

1. lock_branch: true is set on main. Worth confirming that is deliberate.
2. enforce_admins: false means three of the four collaborators bypass the gate entirely. Required checks are real for the fourth and advisory for everyone else until that flips.