Check that every documented status is produced by a test

[Deval123]

What changed

OpenApiSpecIT checked paths and methods in both directions, but statuses in only one. A test producing an undocumented status failed. A documented status that no test produced passed silently. Now both directions are checked.

• assertStatusDocumented, the method every per-status test reaches, also records the (method, path, status) it saw.
• After the class, the documented set must equal what was recorded plus PROVED_ELSEWHERE. The documented set is every key under paths.<path>.<method>.responses in docs/openapi.yaml. PROVED_ELSEWHERE is a named map from each status this context cannot produce to the class that proves it. Today it holds the two 501s, proved in FeatureNotOfferedApiIT.
• Three failures are reported separately: a documented status nobody produces, a stale PROVED_ELSEWHERE entry, and a response key that isn't a status code (such as default). No single test can produce the last kind, so the check fails on it rather than dropping it. The spec has none today.
• WholeClassRun, in server/src/test/java/dev/nkap/server/support/, is a general JUnit extension with nothing OpenAPI-specific in it. It counts the tests that ran against the @Test methods declared, including inherited ones. It runs the check only when the whole class ran and passed, and logs one line when it skips. It refuses a class with parameterized, repeated or dynamic tests rather than miscounting them.
• The javadoc sentence calling coverage "a convention maintained by review" is replaced by what the mechanism is and what it doesn't cover.

Seen failing

Each change below was reverted afterwards, and each run was the whole class, 44 tests:

1. A status no test produces: I added "418" to GET /payments/{reference} in the spec. The check failed with documented in docs/openapi.yaml, but no test in OpenApiSpecIT produces it … ["GET /payments/{reference} 418"].
2. A missing per-status test: I removed @Test from get_payment_404. The check failed with ["GET /payments/{reference} 404"].
3. A missing exception: I removed GET /balance 501 from PROVED_ELSEWHERE. The check failed with ["GET /balance 501"] as documented but unproduced.
4. A stale exception: I removed 501 from GET /balance in the spec and kept the PROVED_ELSEWHERE entry. The check failed with named in PROVED_ELSEWHERE, but docs/openapi.yaml no longer documents it: a stale exception … ["GET /balance 501"]. It was reported only as stale.
5. One test alone: I ran OpenApiSpecIT#get_payment_404 by itself. It passed (Tests run: 1), and this line was logged: OpenApiSpecIT: 1 of 44 tests ran, a partial run, so per-status coverage of docs/openapi.yaml was not checked. Run the whole class to check it.

WholeClassRunTest also covers both branches directly: a whole run runs the check, a partial run skips it and logs, a failed test skips it and logs, a failing check fails the class, and an uncountable test is refused.

Measured and assumed

Measured:

• The four tests that call assertStatusDocumented directly are exactly the ones the plan named.
• One more test asserts a status outside the funnel: post_callback_with_reference_settles asserts 202. It's a behaviour test, and that status is recorded by post_callback_with_reference_202, so it leaves no gap.
• The spec documents 43 numeric statuses and no other keys. On a whole run, all 41 that can be produced here are recorded, so no per-status test is missing.
• mvn -B clean verify passes on the committed tree, with the check run in full.

Assumed: Surefire's default exclusion of nested classes keeps WholeClassRunTest's fixture classes from running as tests. The report confirms it: 6 tests run, none from the fixtures.

Not measured: The JUnit Platform test kit isn't on the classpath, so rather than add a dependency, the extension's tests call its callbacks directly with a mocked ExtensionContext. The real one-test run is item 5 above.

One addition to the plan: it said "equal counts: the assertion runs". This also skips the check, with its own logged line, when the whole class ran but a test failed. A failed test may stop before recording its status, so the check would report a gap that is only that failure's echo, and the build is already red.

Closes #230