Scalar rescue preserves baseline recovery

Protocol frozen at cb8c265c before evaluation on 2026-10-30. See protocol, plan, complete results and schema.

Control	Joint baseline	Fixed-pose scalar rescue
Perturbed pose+bias recovery	33/36	33/36
Truth-start pose+bias recovery	11/18	17/18
Numerically terminated baseline joint fits	50/54	50/54

All 54 baseline fits reran exactly equal to their archived records. Every output pose is exactly the baseline pose. Six residual probes triggered scalar search; all six candidates were adopted. These are the six negative-bias truth starts, but neither bias labels nor seed labels enter the decision. Six recovery gains, zero regressions. Corrected bias errors range .000218–.006862 m. All adopted costs decrease; all other solutions retain the baseline. Actual optimizer work, including joint finite differences, probes and scalar evaluations, is 11–639 calls per case, below the fixed 800 limit. Runtime including diagnostics was 17.51 seconds on this environment; this is not a controlled speed comparison.

The remaining four failures are unchanged: near_foreground positive-bias positive seed for both populations, and staggered all-return zero-bias positive seed and truth start. They retain the original joint iteration-limit failure. Scalar termination cannot certify convergence of an unsuccessful pose optimization.

An adopted result combines the original joint termination with a successful fixed-pose scalar optimization. It does not establish joint stationarity at the corrected bias. Result status/nfev/njev refer to the original joint stage; baseline_fit preserves it, rescue records the scalar stage, and actual_residual_calls counts both. Support is unchanged because pose is fixed; robust sensitivity is recomputed before adoption. Cost descent alone is not an accuracy certificate.

This supports a limited bias rescue mechanism, not production adoption or SOTA: the sampler still knows the synthetic scene. Nine focused tests pass, including fallback on unaccepted baseline, insufficient budget, no improvement and candidate degeneracy. Report/schema, all frozen hashes, exact baseline equality, pose preservation, budgets and cost monotonicity were checked.

Next prioritize the known deployability gap: freeze an image-only depth-sampling comparison against the existing bilinear baseline. Keep this oracle result as an upper-bound diagnostic; do not spend further iterations tuning these four oracle failures. Include all bias signs, both seed directions, truth starts, occlusion populations and planar degeneracy controls before any real-data adoption claim. Separate the sampling comparison from scalar rescue to attribute changes clearly.