Weak/jet serial stills are often geometry-limited: a few-px beam error or a mm-scale detector
distance error (unreachable per-image) fails many frames, but is well-determined jointly from the
strong frames. Mirror the rotation two-pass: an index-only first pass over a spread sample collects
each indexed frame's spots + assigned HKL + orientation; the strongest ~N (default 200) feed one
Ceres bundle adjustment; the refined geometry is applied and the main pass re-indexes + integrates
+ merges every frame from scratch.
GeometryRefiner (reusing the extracted XtalResidual - the RecipToDetector geometry residual pulled
out of XtalOptimizer, behaviour-preserving): one problem with SHARED beam(2)/distance(1)/cell-length(3)
blocks + a PER-FRAME orientation(3) block, robust Cauchy loss, a cell-length regularizer anchoring the
known cell to break the low-resolution distance<->cell-scale degeneracy, DENSE_SCHUR eliminating the
per-frame orientations, and a 3-round HKL-reassignment / tolerance-tightening loop. Tilt is not refined
(gauge-coupled, zero gain). Opt-in via --refine-geometry[=N]; stills only (rotation untouched).
Validated: KR2 7.58% -> 21.85% (matches CrystFEL's 21.5%; a real ~1.4mm distance error + ~3px beam),
OCP 2.92% -> 4.19% (~3px beam). OFF runs are bit-identical to baseline (XtalResidual extraction
non-regressing; rotation lyso_ref de-novo ISa 17.3 unchanged).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
leonarski_fandClaude Opus 4.8
2f5ed411e2