florisvb / PyNumDiff / 33700351915

94%

push

github

Drop the solver knobs and tidy lineardiff's convex fit

The `solver` parameter is gone from `lineardiff`, `tvrdiff`, and `smooth_acceleration`: the first pins CLARABEL, the other two let CVXPY dispatch, which beats any fixed choice. SolverError is no longer caught anywhere, so a real solver failure surfaces rather than being laundered into a merely-bad score, and `lineardiff`'s unregularized `lstsq` fallback is replaced by a raised LinAlgError, since it never fired and solved a different problem than the caller asked for when it did. `robust_data_scale` computes the MAD directly instead of through scipy, and `estimate_integration_constant` gained an analytic gradient for its Huber fit and now flattens to one canonical shape rather than reshaping at every use; together those are worth about 15% of `lineardiff`'s runtime. Its search space now tries both kernels, and one `smooth_acceleration` error bound is tightened to what it actually achieves.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

24 of 26 new or added lines in 4 files covered. (92.31%)

813 of 866 relevant lines covered (93.88%)

0.94 hits per line