Independent optimization audit: PASS

Algorithm: Same coverage objective, retained-tail enumeration, greedy additions, feasible improving exchanges, deterministic exact ties/order, validations and single original independent check. Changes only cache F-layout candidate columns and reuse buffers.
The original 3/4 proposal guarantee retains its original sampled-objective scope; it does not extend to final checked output or realized/field gains.

Freeze: All three implementation hashes, original kernel/API/archive hashes, ten input hashes and development/heldout evidence hashes match. File timestamps and fold0-before-heldout chronology are consistent. No QUERY_TRUTH read.
Full evidence: 3276 parents / 6552 degraded / 3276 complete. Reviewed full nested receipt assertions; independently checked every stored output/proposal, core/quota/uniqueness/order, check outcome, gain/LCB, archive identity, axes and seeds.
Bounded independent runs: 9 degraded queries including one improving exchange and one rollback; 4 complete identities; 180 synthetic cases over K0..4, C/F/strided arrays, exact/ULP ties; 8 invalid-input cases; complete/K0 paths ignore scenario objects. All exact serialized receipts match.

New same-run serial matched benchmark: 80 parents, 160 degraded queries, 3 repeats, 480 original/optimized pairs. All row keys/order/receipts match; CSV statistics independently recalculated.
Original mean/median/P95: 19.063478750 / 17.815500000 / 27.643385000 ms.
Optimized mean/median/P95: 5.534065417 / 5.258400000 / 7.572235000 ms.
Ratio of same-run means: 3.444751248x.
Public selector validation, complete receipts and the original independent check are included. Shared scenario generation, candidate generation, inference, I/O and equality assertions are excluded.
Historical 24.71 ms is a separate all-method run. It was not used for this speedup. Do not mix old baseline times with the optimized time as a matched comparison.

Limits: Exact equality applies to tested archived inputs and numerical environment. Full nested replay was performed by the reviewed frozen verifier; audit independently repeated bounded cases. Shared scenario-generation mean is reported only, since raw generation durations were not retained. No accuracy or end-to-end speedup claim.
