Add one-shot JSON-repair retry to filter_score on parse_error

After the parseJsonLoose failure check at lines 142-152, before recording the parse_error move, attempt one repair call: re-feed the raw response back to the same model with a stripped 'extract the scores object as valid JSON, no prose' instruction at low temperature and a 30s timeout. If the second-pass parse yields valid {scores, justifications}, treat as if first call succeeded and continue down the success path at line 154. If it fails, fall through to existing parse_error recordMove. Gate behind an env flag (REPAIR_RETRY_DISABLE=1) for instant rollback. Touch both the high-reviewer and low-reviewer branches symmetrically — only the one that failed parse gets repaired.

Across the next 50 filter_score invocations, parse_error rate drops by >=60% and the corresponding cost-without-phase-advance burn drops proportionally. Repair calls cost ~10-20% of an original call.

Log repair attempts for 50 sequential filter_scores. If parse_error rate is unchanged or repair-pass success rate <40%, the harness is not solving the parse pattern observed and should be removed.

Rejected by filter_score. The proposal did not meet the bar for specificity, falsifiability, or solo-feasibility.