The rotation geometry post-refine committed the detector distance and cell
scale on any cross-validated improvement within a loose 3 % / 2 % bound. On a
multi-lattice crystal the second lattice's spots bias every cross-validation
fold the same way, so the "it improved" test cannot catch it and a spurious
distance<->cell pair (their radial degeneracy) is committed far off the true
geometry, losing reflections in the re-integration.
Tighten the commit bound to < 1 % (a calibrated header needs < ~0.6 %). The
move SIZE discriminates a genuine header correction from the multi-lattice
failure far better than the fit residual, which genuinely marginal (noisy or
ice-ring) data shares with the multi-lattice case - so a real ~0.3 % correction
is kept while the ~1.7 % multi-lattice one is rejected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A first pass integrates and post-refines the detector geometry from the observed
diffraction, applies it to experiment_, and a second pass re-indexes and
re-integrates with it - only the second pass is written, so the _process.h5 / mtz
(and the ice-ring flags, which move with the geometry) reflect the refined values.
Default off; enable with --rotation-post-refine.
PostRefineRotationGeometry refines in two SEPARATE cross-validated steps (a joint
fit of everything is fragile - the same lesson as integration, where refining the
profile width and the scale jointly fails but separately works):
* Step A: cell scale + rotation axis from the observed rocking centroids phi_obs
(a distance-independent excitation residual, so it pins the absolute cell scale
that the positional residual leaves degenerate with distance).
* Step B: detector distance + beam centre from the observed spot positions, with
the cell fixed at step A.
Each step commits only if it lowers a held-out (deterministic split-half) residual,
else that part of the geometry is left at nominal.
Rugnux::Run wraps RunPipeline(write_output, geometry_prepass); the pre-pass runs
the first scale/merge so RotationScaleMerge fits a frame-order-smoothed mosaicity,
which is captured and fed to the second pass's Bragg prediction (via
IndexAndRefine::SetPredictionMosaicityOverride) rather than re-derived per image.
observed_x/y were plumbed through the profile integrator earlier for the positional
residual. rugnux_vs_xds.py gains --extra-args to sweep the option over the battery.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
leonarski_fandClaude Opus 4.8