Source fd75a63657bc6b1d8cbd26c644bd8fb39cc1a3ec; reviewed harness 11b7bc74e4756bafc2c6c450628d10c6a62f8f39. Native affine fixed-plan disturbance propagation and convex vertex worst-case objective. MSVC Release, HiGHS/Qhull disabled. All assurances are numerical.
| Fixture | Expected or sourced | Observed | Evidence scope |
|---|---|---|---|
| L19 | (x−1)² on [0,2]: midpoint 0; endpoint worst 1 | Public evaluations [1, 0, 1]; global V-row worst 1 at x=0 | Authored compact V set and numerically admitted convex objective |
| Integrator, feasible | Four terminal corner candidates [-2, 0, 0, 2]; worst x²=4 | Four candidates; worst 4 | Two independent compact disturbances; convex H safety |
| Integrator, violation | Stage-one successor −1 outside [-0.5,0.5] | Robust violation at stage 1, successor −1; no tube | Whole-set containment, without clipping |
| B08 rounded | Independent saved discrete reference 67.150601101407, three ties at 1e−8 | 67.150601101407545; 625 plans × 256 disturbance histories, three ties | Displayed rounded matrices and finite action catalogue only |
No numerical residual vector is exposed by these two public APIs. The published 67.12, original unrounded MATLAB arrays, nonlinear trajectories, historical runtime and validated rounding are not reproduced. The B08 finite-plan reference does not establish feedback optimality or a complete optimal family.
One excluded warmup then seven measured rounds per fixture; zero process or literal-oracle failures. Durations are batch milliseconds, not single-call latencies. Each timed interval includes fixture construction, public API execution, literal checks and JSON result construction. It excludes process startup, serialization, persistence, plotting and rendering. Different batch sizes make fixture rows unsuitable as a speed comparison.
| Fixture | Calls/batch | Successful rounds | Median ms/batch | Range ms/batch |
|---|---|---|---|---|
| l19 | 10000 | 7/7 | 45.804 | 45.123–57.474 |
| integrator_feasible | 5000 | 7/7 | 97.210 | 96.531–108.355 |
| integrator_violation | 5000 | 7/7 | 76.544 | 73.668–78.146 |
| b08_rounded | 1 | 7/7 | 91.284 | 90.217–94.758 |
The L19 batch is short (about 45 ms); its per-call time is not inferred. Host load observations were 8.8%, 1.3%, and 1.4% around the runs. The separate profile recorded process CPU and resident memory, but did not identify a source-level limiter. Performance remains exploratory and inconclusive; no comparator or speedup is claimed.
The first profile collection blocked on an unread pipe; its output bytes were unavailable and no accepted trial had begun. A second complete run failed a provisional 50 ms batch gate; its raw output is retained but excluded. The final run used a 20 ms admission floor after another fresh clean Release rebuild.