Image-only scalar rescue recovers negative-bias truth starts

Frozen at 0b0fa0bb before evaluation on 2026-10-30. See protocol, plan, all 63 results and schema.

Only the residual/sensitivity/joint-fit imports change from the frozen scalar rescue algorithm. Its function AST is identical. The image-only sampler, joint optimizer, objective, thresholds and support/rank gates remain unchanged. All 63 baseline fits rerun exactly equal to their archived records; all returned poses are exactly the baseline poses. No oracle scene information is sampled.

Sphere controls	Image-only sampler	Sampler + scalar rescue
Perturbed pose+bias recovery	32/36	32/36
Truth-start pose+bias recovery	10/18	16/18
Median absolute bias error, all 54	.002071 m	.001974 m
Baseline joint numerical termination	54/54	54/54

Six gains and zero regressions relative to the image-only baseline. All six negative-bias truth starts are recovered; this includes the two staggered cases that regressed when moving from bilinear to gated sampling. Their corrected bias errors are .001494–.005005 m. All nine parallel-plane controls remain rejected as degenerate and never enter rescue.

Eight residual triggers lead to eight adopted corrections. Two are near_foreground positive-bias positive-seed cases: cost and bias error improve, but error remains .148018/.159259 m, so both still fail. Four other failures are unchanged: original positive-bias positive seeds (.164649/.164893 m error) and original positive-bias truth starts (.050904/.051026 m error). Do not round the latter into success or adjust the .05 m gate. Pose-only accuracy passes for all 54 sphere cases; numerical termination does not certify accurate bias estimation.

Actual optimizer work is 8–215 calls, including the baseline finite differences, probes and scalar trials, below the unchanged 800 limit. Diagnostics are accounted separately as in the baseline. Runtime including diagnostics is 21.47 seconds; this is not a controlled speed comparison. Adopted corrections strictly reduce the measured cost; all fallbacks retain baseline results. The nested baseline_fit and rescue audit preserve both stages. Inherited status/nfev/njev describe the joint stage; adopted results have staged joint/scalar termination, not demonstrated joint stationarity after the bias update.

Ten focused tests, repository Ruff and strict documentation build pass. Checks also verify schema, all frozen source hashes, the 63-case denominator, exact baseline reproduction, unchanged poses, nonincreasing cost and hard call budget.

This closes the specific negative-bias truth-start regression on these synthetic controls. It does not establish production readiness or SOTA; six sphere failures remain and the data are spent. Core/defaults are unchanged. Next prioritize a pre-registered transfer check on the existing real-data development recordings, with the fixed image-only sampler/rescue and a comparable bilinear metric baseline. Report all drives and both perturbation signs, truth-start drift, rejection and runtime. Real depth has no known injected-bias truth, so do not reuse synthetic bias-recovery claims there. Do not tune these six residual synthetic failures or consume fresh holdouts before freezing the real-data comparison.