Parity compares exit codes, so two implementations wrong the same way are green #148

test_verdict_agrees_with_the_reference_harness runs both implementations over the corpus and asserts the overall verdicts match. That is one bit per run.

What it cannot see

• A shared mistake. Both implementations pass, both are wrong, the test is green. There is no oracle in the comparison; each side is only checked against the other.
• Compensating differences. One implementation misses a violation while the other reports a spurious one. Same exit code, opposite errors.
• Which case disagreed. When it does fail, the assertion prints our failures and the tail of the reference's stdout, and a human diffs two differently-shaped reports by eye.

Shape

Compare per case rather than per run: each implementation emits a machine-readable report of case to outcome, and the suite diffs them against the expectations the corpus declares.

Two things fall out. The comparison stops being pairwise, so a third implementation costs nothing here - it runs the same suite and produces the same report shape, and no implementation needs to know about any other. And this package stops being both a participant and the judge, which is the part that does not survive a third port.

Depends on OO-LD/oold-schema#158, which is where the expectations have to exist before anything can be diffed against them. Until then the exit-code comparison is the honest maximum and should stay.