Accelerate CRISPRWorks Fit 8.59× with a compiled EM engine

[dnncha]

Fit's NumPy accelerator still spent most fitting time in Python's EM loop. Add an optional compiled C engine that reuses iteration storage and omits exact zero design entries for finite inputs, retaining a dense path for nonfinite values. Preserve MAGeCK2's statistical conventions, support explicit native/NumPy engine selection, and record the resolved kernel in provenance. Bump the source alpha to 0.1.0a2.

Three paired runs per engine on a public HAP1 cohort of 17,445 gene labels and 69,780 guides measured median wall times of 332.452 s MAGeCK2, 131.060 s previous NumPy engine and 38.681 s native engine: 8.59× versus reference and 3.39× versus NumPy. All nine printed gene summaries were byte-identical; permutation p-values/FDR matched exactly. Maximum beta difference was 2.49e-14. Timings include startup and diagnostics; the cohort selects complete four-guide labels before normalization. Publish full machine-readable records and reproducibility instructions in benchmarks/NATIVE.md.

The unfiltered 71,090-guide, 18,056-label table also matched reference outputs in a separate cross-environment numerical check; no full-table timing ratio is claimed. Linux 3.10/3.11/3.13, macOS 3.12, installed wheels, spawned workers, batch screens and the container passed CI at engine commit 2938b8e. Local native tests passed AddressSanitizer and undefined-behavior checks. Add an explicit native singular-solve regression check in the evidence commit.

This is a source-only alpha and a measured compatibility/performance result. Biological superiority, universal performance, or overall SoTA versus Chronos/JACKS is not claimed.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.