test: cover the exit-code wiring of the corpus checks

[yakisoba0728]

The gap was not where it looked

The premise going in was that validate_corpus.py, check_decode_errors_corpus.py and check_metrics_baseline.py print SKIP and exit 0 without a corpus, the trap CONTRIBUTING.md warns about. Measured: all three exit 2 with a named reason. They already follow the doctrine. (The warning does apply to check_export_baseline.py and check_corpus_baseline.py, which skip when their single pinned replay is absent.)

The real gap is that the layer CI and developers actually trust — main() reading a verdict and choosing an exit code — had no tests at all. The helpers underneath (problems(), dead_counters(), reconcile(), invariants(), compare()) are well covered. The wiring between helper and exit code was not, and this repo has been burned there twice: a counter that vanished while the run printed WARNING and exited 0, and a run that decoded nothing while printing OK.

What this adds

22 tests across the three tools/tests/ files, following the existing convention for faking an executable. Each mutation reproduces a historical failure:

mutation	what happens without the test
disable the invariant gate in check_metrics_baseline	a run with rounds=0 and score={} is pinned into baseline.json, exit 0 — the exact shape of the 13.02 RoundResults regression
disable dead_counters in check_decode_errors_corpus	prints "Decode errors: 0 ... over 0 decoded rows", exit 0 — the bug its own docstring cites
ignore problems() in validate_corpus	prints "OK: 0/1 replays validated", exit 0

check_metrics_baseline's extract() also had no coverage of the nested valplay metrics shape it reads — a cross-repo seam that moves without warning. Five tests pin it.

Verified against the real 215-replay corpus and a locally built vrfkit before writing any of this, so the scripts' behaviour is measured rather than assumed.

Also: a pre-existing CI failure

check_docs.py --fast was already failing on main — three lines claim 117 ascii-checked Rust files where check_ascii.py reports 118. The docs step runs before the tools test step in python-checks, so without this fix the new tests would never have run in CI.

Not built, and why

• A miniature corpus. The checksum_export fixture works because it is a structured JSON manifest, not raw replay bytes. These scripts feed real .vrf bytes to the real binary. crates/vrf-container/tests/corpus.rs already has a hand-built fixture and scopes itself to container smoke-testing — it never touches the replication payload inside a content block, and hand-authoring that is a project of its own. validate_corpus.py's claim is also a distribution claim across the corpus, which one hand-made replay cannot support.
• A workflow_dispatch job with an out-of-band corpus. Cross-repo checkout needs a PAT, LFS bandwidth bills the other account, and above all a manual job gates nothing on an ordinary PR.
• A "make the SKIP visible" job. There is no silent SKIP to surface, per the measurement above.

Verification

python -W error -m unittest discover -s tools/tests -p "test_*.py"   503 passed (was 481)
python tools/check_docs.py                                           exit 0, rust 563 / tools 503

No workflow change: python-checks already discovers tools/tests.

Handoff

metrics_builds.json shows drift on build 13.01 (kast_rounds 117→116, trade_kills 35→33), consistent with valplay's recent self-damage attribution fix. CONTRIBUTING.md requires explaining each changed line before --update, and that is a domain judgement — left alone.