Third-judge tiebreaker harness on high-IQR filter_score runs

After the 3rd filter_score run completes and median/IQR are computed, check if iqrTotal ≥ 1.5 (high disagreement across the 3 GPT-vs-Gemini midpoint runs). If so, invoke a single tiebreaker call to a third model (Codex via llm.callCodex — already used by meta_engine judging). The tiebreaker scores the same 5 filters once, and its run total is averaged with the median to produce a tiebroken_median which replaces median in the composite formula. Tiebreaker is logged as a separate move row (move_type='filter_score_tiebreak') for cost auditability. Threshold and 'on/off' controlled by config flag.

Currently ~10-20% of completed scoring sequences end with IQR ≥ 1.5 (visible in `filter_score_iqr` histogram). For these high-disagreement candidates the tiebroken_median will shift composite by ≥0.5 points in ≥50% of cases, changing the top-of-queue ordering. Promotion/kill decisions on candidates currently sitting near the score floor should flip on roughly 1 in 8 high-IQR runs.

Run on next 15 candidates that hit the IQR ≥ 1.5 condition. If Codex tiebreak total is within ±0.5 of the median in >12 of 15 cases, the disagreement was symmetric noise and Codex adds no signal — harness is wasteful. Disable and refund the cost budget.

Rejected at the council verdict. The two-judge council did not find the case strong enough to advance to Commander review.