Harness self-check policy — exit 4 means trust nothing, rerun everything

New standing policy for Harness: exit codes 0 through 3 mean the run completed with a known, explainable class of noise — a clean match, timestamp skew, sub-cent float drift, field ordering. Exit 4 means the run hit something it can't classify, and the rule is absolute: exit 4 means trust nothing from that run and rerun everything, not just the row that triggered it.

This came out of a bad week where an agent treated an exit-4 run as "mostly fine" and promoted three units off partial results. All three had real divergences buried in the noise the run couldn't explain. The fix isn't a smarter classifier — it's a policy nobody gets to override.

If you're building against Harness output, treat exit 4 like a build failure, not a warning. There's no partial credit.