[abhishekgahlot2]

• QA gate: unknown attack strategies are rejected before any attack runs (422 over the API). A misspelt one ran a no-op and read as robust.
• QA gate: an audit drops a changed verifier's stored result before its retest, so an inconclusive retest no longer leaves the old result standing.
• Invalid case scores and NaN or infinite reward components no longer become a trainable 0 under a valid overall score. The verifier wrapper checks them like the overall score: invalid after the solution ran is a judged 0 (solution_error); invalid from the verifier alone is invalid_score with no verdict. The host rejects them too (InvalidScoreError).
• OpenRLHF: a sample whose verifier failed or returned an invalid result gets infra_failure_score instead of 0. Judged model failures and runs still going at the request deadline still score 0.
• Harbor export refuses snapshot-backed environments instead of exporting them on a plain image.
• RL trace rows carry initial_observation and initial_screenshot.

Each fix has a regression test that fails on main.