Serial stills currently treat every reflection as a full (partiality hardcoded to 1).
Add an opt-in Gaussian excitation-error partiality set at prediction time (CPU + CUDA):
p = exp(-dist_ewald^2 / (2*sigma^2)), sigma^2 = profile_radius^2 + (bandwidth_sigma*|recip_z|)^2,
with sigma = the per-image profile radius (ewald_dist_cutoff/2), so an edge-of-acceptance
reflection keeps p ~ exp(-2). Off by default; the merge weight (~p^2) then down-weights
far-from-Ewald partials instead of trusting them as fulls.
Validated: helps medium/strong stills (LOV R-free 0.336->0.329, lyso8 0.433->0.410, lowers
the systematic error-model b in both) but HURTS weak OCP (dividing by a small, uncertain p
amplifies orientation error -> high-res noise, resolution collapse), so it is left opt-in.
A static forward p explains only ~6-10% of the partiality scatter; the full win needs per-image
post-refinement (future work), for which this is the prediction-side groundwork.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
leonarski_fandClaude Opus 4.8
fb399480e5