Masked/error pixels carry the int type minimum and saturated the maximum,
but the lossy codec can nudge them inward by one (masked observed as
INT32_MIN+1 in the decompressed data). The integrators only checked the
exact extremes, so a shifted sentinel in a reflection's background ring was
treated as a valid pixel -> garbage background and intensity for any
reflection whose box clips a module gap. Reject the +/-1 band too (real
calibrated counts never approach the type extremes). Neutral on the
well-centred lyso test set; a correctness fix for gap-clipping reflections.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reference dataset (a): LoadReferenceMtz adds column selection + cell/SG/resolution +
a data-vs-reference consistency check; jfjoch_process/jfjoch_scale gain
--reference-column; the viewer gets a Reference section in the MX settings dock
(worker-owned, independent of the loaded dataset) that flows into reprocessing jobs.
3D combine (-P rot3d): Combine3D weight-sums a reflection's per-frame partials into one
counting-limited full before merging (orthogonal ScalingSettings::combine_3d flag, not a
partiality model), with a de-biased Poisson variance. Crystal 2: ISa 1.7->8.4, R-meas
~67%->18.9%, intensities unchanged (CCref held).
Quality metrics (b): R-meas (Diederichs-Karplus) + redundancy columns in MergeStats; ISa
logged. jfjoch fulls 18.9% vs XDS 4.5% (same ASU/run).
Profile-fit integrator (experimental): ProfileIntegrate2D (--integrator gaussian|empirical)
is a reference-free, rot3d-compatible profile-fit extraction (the decomposed PixelRefine
intensity step). Gaussian: R-meas 18.9->14.6%, ISa ->9.5. Anisotropy/per-region add nothing
(the discriminating info is in the discarded rocking direction). See NEXTGEN_INTEGRATOR.md.
--dump-observations exports the unmerged fulls for XDS comparison.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
leonarski_f