Offline direct-disparity pose landscape

Pre-registered 2026-10-30. Diagnose all 171 archived direct-disparity fits without new calibration fits, changed gates or new data. Reconstruct exact depth arrays, verify hashes, and probe the original objective. Preserve clean/noisy, both seeds, all biases, signs, populations and nine plane controls.

For each case, interpolate from reference pose to archived fitted pose at eleven fractions 0,.1,...,1. Use linear camera-frame translation and shortest-path rotation SLERP. Evaluate two tracks: fixed injected metric bias, and fixed fitted metric bias. Reference and injected bias are offline diagnostic information only; no truth-dependent initialization/gate is proposed. No bias optimization occurs. Record total cost, support count and cost restricted to the intersection of supported point indices over the whole track. Reproduce fitted-bias endpoint cost against the archive. Keep all points and original unsupported penalties.

Report reference-versus-fit preference for each noise realization and clean data, separating perturbations from reference starts. Also report sampled interior minima and whether any interior sample falls below both endpoints. A favorable reference endpoint does not establish a global optimum, local stationarity or an optimizer's internal failure cause. Eleven samples cannot rule out narrow barriers or off-path minima. Fixed injected bias under noisy metric offsets is not a maximum-likelihood nuisance estimate. Plane probes remain diagnostic only. No revisions to recovery counts, production adoption or SOTA claim. Freeze code, protocol and archived result hashes before running; preserve partial output on failure. Diagnostics have their own pass count, not a new optimizer budget.