Reuse Fourier preparation in four-case SSB review

Historical: describes the code as of 2026-10-22; its fits refined with the finite-difference Newton, since replaced by exact derivatives, so its fit timings are not current. See the current fit methods.

The rotation/defocus comparison now prepares one temporary full-field session, then reuses its source Fourier data for all four cases. Changing the rotation refreshes the geometry. Each signed-defocus search still starts from zero aberrations and has its own 200 trials plus Newton refinement. The caller’s rotation, calibration and fitted optics are preserved.

Matched CUDA evidence

The private experiment group 1022-rotation-reuse compares frozen source db305309 with the candidate on a physical Blackwell Max-Q GPU. All valid bright-field pixels, detector dimensions, precision and estimators are retained. Fitting uses native scan sampling; displayed reconstruction uses Auto.

Median wall time for the complete four-case tilt/depth comparison on 512-square scans, excluding acquisition loading:

Acquisition	BF pixels	Auto	Before	Reuse
A	8,807	3x	4.551 s	4.190 s
B	8,930	4x	7.493 s	7.140 s
C	101	1x	0.249 s	0.218 s

Three matched repetitions provide initial timing evidence, not robust p95 latency. Acquisitions A and B retain 192-square detectors; C retains a 12-square detector and already has finer scan sampling than its aperture needs. Do not compare these detector workloads as equivalent.

Numerical acceptance

All three acquisitions passed at 256- and 512-square scan sizes, with ordinary and tilt/depth-corrected SSB tested separately. Complete native and Auto phase arrays and fitted coefficient records matched the frozen implementation exactly. A subsequent run through the changed public API also matched every frozen phase array and fit record. A contract test verifies independent starts, unchanged trial budgets, Auto output and preservation of the caller’s state.

This change reduces redundant preparation; it does not omit any search or establish the physical rotation. The result remains unresolved without an independent focus or rotation constraint. MPS hardware acceptance remains separate from this CUDA study. No release or default estimator change is made.

Consuming notebook acceptance

The changed API also passed in a real Jupyter/ShowPtycho workflow on acquisitions A and B at full 512-square scans. The notebook loaded the acquisition, ran the 200-trial primary fit with Newton and tilt correction, displayed Auto phase and FFT, and responded to a pointer-driven C10 adjustment. The subsequent four-case comparison matched all eight frozen native/Auto phase arrays per acquisition, left the parent session unchanged, and rendered one comparison figure. The browser reported an NVIDIA Blackwell adapter; sessions were closed after testing. This is functional workflow acceptance, not an interactive frame-rate benchmark. The private evidence group is 1022-consumer-notebook-v6.

Additional acquisition coverage

A fourth acquisition with 9,298 valid BF pixels and Auto 3x output also passed complete native/Auto phase and fit-record equality at both scan sizes, with ordinary and tilt/depth SSB tested separately. Six matched comparison sessions per variant used the same Blackwell workstation GPU. These sessions reused one loaded acquisition and are not fresh-process loading measurements. The private evidence group is 1022-rotation-reuse.