The configured settings study compares ordinary Native/BAB with selected explicit routes in the same solver build. Eligible capacity models retain DP; linear families can enable checked LP, cover cuts and bounded reliability probes. Native global routes remain active for routing and queens. Every optional mechanism reports actual work. An exploratory pilot selects a frozen family policy before adaptive timing; these reused seeded families do not constitute an independent holdout.

The adaptive boundary study searches each native version independently under ten-second limits. Expansion stops after a failed size, followed by bracket search and confirmation. These are sampled boundaries, not absolute solver limits. Original witnesses are checked independently; exact optimality is reported by the backend without a separate reference oracle. Earlier fixed studies remain separate records.

The seven-category study records 132 runs from 33 frozen cases under ten-second limits. Knapsack is grouped across both distributions and both repetitions; its largest tested proved size rises from 32 to 384 model variables. Six other category maxima are unchanged. Routing and queens report actual integer model variables separately from city/queen counts. The single-variant 512-item stress case remains visible. No general maximum, speed or commercial-parity claim follows.

The capacity scaling study measures 13 frozen cases with two paired repetitions and ten-second limits. The largest tested knapsack size proved in both repetitions rises from 32 to 384 Binary variables in two fixed distributions; assignment remains at

The 512-variable table-overflow point remains in the census and times out for both versions. These are sampled sizes, not general solver limits or speed ratios. All 52 original-model witnesses pass independent checks.

The algorithm checkpoint compares the bounded binary capacity optimization with the immediate prior implementation using the same driver and checked runtime-library cohorts. The separate before-project and eighth-to-ninth studies are historical. None is a CPLEX/Gurobi comparison or a rerun of the entire original corpus. Existing CP tests exercise original native functionality; Optimize tests separately exercise the additive facade and its backends.

The expanded FAST gate passes 35 required cases, including selected-basis sensitivity and an actual bounded neighborhood improvement. Correctness configurations overlap and their counts must not be added as unique coverage. The before-project panel compares preserved original experimental binaries with the ninth checkpoint, including its HiGHS binary-model route; build and loading costs are part of that measurement. The separate eighth-to-ninth comparison uses preserved executables and common workloads; one attempted extra panel exceeded the unchanged cp_distinct deadline and remains disclosed. No limit was relaxed to obtain a pass.

See the benchmark report for measured values, charts, cohort identities, retained failures and limitations. That report is included as a native documentation page. Its portable source is docs/solver-parity/BENCHMARKS.md; the checkpoint record and raw artifact paths are retained with the source tree. Performance evidence applies only to the stated workloads, builds, machine and budgets. The broader solver-parity roadmap remains incomplete and further feature implementation is on hold.