Deterministic empirical-refit comparison

Six complete CLI runs retain three alternating pairs, each with twenty seeded temporal-block refits. The prior runtime uses one process; the current runtime uses four. Whole-process medians are 29.935852 s and 12.643885 s (2.367615x). Every run retains 6,224 solver trace entries. All 120 refits and six uncertainty results match exactly after the declared execution metadata normalization.

The benchmark report describes usage, numerical behavior and limits. protocol.json binds the complete source capsules, all three input identities, launcher/freezer/schema and dependencies. fixture-builder.py.txt is the exact original builder named by the fixture protocol, with nine raw KITTI hashes and deterministic point selection. Coordinates, generated correspondence/problem inputs and nested Git metadata remain external. The control uses public KITTI geometry with generated image correspondences. It establishes execution parity, not real-image accuracy or uncertainty coverage. All uncertainty policies remain inconclusive.

Each run retains its twenty refit YAML files, final uncertainty artifact, stdout/stderr and imported-source paths. The four process initialization receipts in each current run also bind child imports to the source capsule. The launcher only observes imports and invokes the regular CLI. Both source trees had real Git identities backed by the local object store; normal Git provenance was retained during timing. No numerical operation was patched.

All attempts are retained; no timing warmup is removed. Both variants use one BLAS thread per process on an unreserved host. The comparison includes process startup, input validation, twenty refits, diagnostics, YAML output and pool shutdown. Separate untimed checks validate every output and child hash before normalizing timestamps, source commits, declared output paths, the worker flag and verified child-output hashes. No numerical field is normalized.

initialization-v1.log records a Git compatibility failure before any timed worker. measurement.log records all six successful v2 workers and the later summary failure on an omitted optional reference. postprocessing.json and postprocessing.log bind the corrected summary to the unchanged measurements; launcher.py.txt is the measured launcher, while auditor.py.txt handles that omitted field and additional audit bindings. No timing was repeated or dropped.

validation/ retains regression sources/logs, schema parity and a receipt binding the final runtime to the measured candidate. index.json binds every other file by SHA-256, preserving bytes across checkouts. Audit with ordinary CPU dependencies, without raw KITTI data or a GPU:

python -m tools.benchmark_empirical_refits --audit docs/assets/empirical_refit_workers