BridgeBank: Predicting Reuse Before Admitting Verified Lean Repairs

A solve-only comparison favors disabling novelty filtering over random reuse ranking on miniF2F solve ($41.4$ versus $41.1$), while the ordering reverses on observed Reuse@12 ($6.2$ versus $9.1$) and fully exposed unused admissions per 100 policy-specific failures (63.4 versus 14.2). This inversion exposes the decision left open by verification: after several repairs type-check, which one, if any, should enter shared retrieval state? BridgeBank trains a top-one-or-none admission score from completed 12-round future-reference cohorts; only typed, retrieval-distinct repairs from diagnosed missing-bridge failures are eligible. Across family-held-out miniF2F and Mathlib-Structured runs capped at 576 A100-hours, it improves Reuse@12 over broad post-failure growth by $+7.1$ [$5.5,8.7$] and $+7.5$ [$5.9,9.1$] points while the solve differences are $+0.5$ [$-0.1,1.1$] and $+0.7$ [$0.0,1.4$]. The resulting libraries contain 520 rather than 1,833 fully exposed but unused miniF2F admissions and 843 rather than 3,488 on Mathlib-Structured. Across recurrence strata, the paired reuse advantage rises from $+2.6$ [$1.0,4.2$] to $+11.5$ [$9.1,13.9$] points. Together, these comparisons identify predicted-reference-guided admission as a route to higher-reuse, lower-clutter libraries at comparable solve in the evaluated regime.