MPS slider correction and interruptible diagnostics

Historical: describes the code as of 2026-10-17; the mean_phase estimator and the phase-variance search are retired. See the current fit methods.

The fine-grid correction evaluates the frequency-dependent tilt shift and aperture bound once per coherent BF group. Previously the helper repeated these expressions for each BF contribution. The expressions, BF order, precision, calibration, output sampling and phase_of_mean estimator are unchanged.

Paired reconstruction measurements

Both private acquisitions contain 512 × 512 scan positions and 192 × 192 detector pixels. A uses all 8807 BF pixels and a 1536² output; B uses all 8930 BF pixels and a 2048² output. Each mode has 20 fresh parameter calls after two warmups, with randomized baseline/candidate order. Times include GPU completion and host phase handoff; they exclude loading, fitting and browser display.

Case	Control / model	Before p50 / p95 / max (ms)	After p50 / p95 / max (ms)
A	C10, tilt/depth	111 / 115 / 115	102 / 105 / 105
A	Tilt, tilt/depth	113 / 120 / 124	104 / 112 / 114
A	C10, ordinary	103 / 105 / 109	92 / 96 / 100
B	C10, tilt/depth	211 / 215 / 215	190 / 197 / 200
B	Tilt, tilt/depth	212 / 216 / 216	189 / 195 / 200
B	C10, ordinary	189 / 199 / 202	168 / 173 / 177

Peak active MLX allocation was approximately 9.72 GB / 10.21 GB. Complete comparison runs took 26.31 / 43.54 seconds. These are active-desktop measurements, not hardware limits or whole-acquisition fit timings.

Cancellation without partial scientific output

SSB.preview(cancel=predicate) is an optional advanced application control. MPS checks it at completed GPU batch boundaries. Cancellation raises concurrent.futures.CancelledError; it returns no partial image or loss. Other backends check before and after computation. Existing calls are unchanged.

For cancellable 512² native loss, the row/column kernels process 256 BF pixels per completion boundary. They carry the original per-pixel phase sums and all eight independent per-thread square sums into the next tile. The original BF group boundaries and final square-sum reduction remain unchanged. Combining the eight square sums earlier would change rounding, so this is deliberately avoided.

A consuming viewer that already holds the completed coherent phase can request native mean_phase with loss and discard that diagnostic phase. This avoids a second coherent image that the viewer would not use. The displayed estimator and native per-BF phase-variance objective remain unchanged.

In paired completed-loss calls, removing the unused image reduced medians from 269 to 191 ms (A) and 280 to 191 ms (B). In eight cancellation trials per case, the maximum remaining time after the cancellation event was 5.33 / 4.57 ms. These kernel-level cancellation times exclude HTTP detection and queueing. Applications must signal cancellation when a new image arrives; browser disconnect notification alone is not a reliable low-latency signal.

Validation and scope

• 188 complete-array comparisons against frozen pre-change GPU source were bit-identical, covering ordinary and tilt/depth models, both rotations, sampling changes, signed C10, tilt, depth and higher-order terms.

• Another 24 full native diagnostic phase/loss comparisons were bit-identical against frozen reconstruction and column-kernel source. Cancellation followed by reconstruction also retained the complete reference phase.

• Sixty focused MPS hardware, contract and privacy tests passed, exercising cancellation and recovery, alternate tile lengths, ordinary/tilted loss and higher-order reconstruction.

• The consuming dashboard verifies C10/tilt movement, diagnostic interruption, full BF, Auto sampling, WebGPU display and recovery of settled loss. All 28 consuming-API arrays matched frozen references exactly. Sixteen real browser interruptions yielded to the next image; maximum queue waits were 8.64 / 5.53 ms. Direct pointer drags and rapid input reached their final requested settings. These are application tests, separate from GPU timings.

• The native harness builds, but the broader frozen fit-trajectory gate remains blocked by an absent local real-data fixture. Native Swift and fitting source are unchanged. This work does not qualify CUDA or establish 60 FPS.

Parallel BF lanes, alternate threadgroups, rectangular support dispatch and individual sideband geometry pruning gave no reliable improvement and were not adopted. Private experiment id: ssb-slider-preemption-20261017.

Integration with ordered coherent reconstruction

The correction reuse also works with the newer ordered BF accumulation and bounded Fourier-support dispatch. It uses their shared transfer function and keeps their group boundaries and summation order. The timings above describe the earlier isolated optimization, not an additional integration speedup.

The combined implementation passed 140 complete phase comparisons against frozen pre-integration GPU source, plus 40 independent native phase/loss and public preview checks on the two full-size acquisitions. These include ordinary and tilted models, both rotation branches, signed defocus, sampling, higher-order terms, and completed calls with cancellation enabled. Every comparison was bit-identical. Private experiment id: ssb-slider-rotation-integration-20261017.