Do not repair recorded harness output to satisfy a validator

When a validator disagrees with a status field an eval harness recorded (e.g. calls that omit a declared required arg but came back result_status ok), resist rewriting the data to agree. Those rows are the evidence that the harness under-validates against its own schema; correcting them yields a clean-looking corpus with the gap still present and no trace of it. Instead make the scorer validate independently of the recorded status so defects cannot hide, then add a test in BOTH directions and mark the failing one xfail with the decision in the reason. The reverse-direction test that passes is the valuable one: proving no valid call was ever marked error establishes the disagreement is one-directional, so the error label stays trustworthy even when ok does not.