1192 Commits

This Branch

This Branch

All Branches

Author	SHA1	Message	Date
leonarski_fandClaude Opus 4.8	fccf9b83e7	RotationScaleMerge: drop profiling + env gates, honour forced mosaicity ... Remove the [rsm] per-stage lap timing and the JFJOCH_RSM_NO_GPU / JFJOCH_RSM_CPU_COMBINE env gates now that the GPU-resident path is the validated default (it runs whenever a GPU is present, with the CPU loops as the bit-parity fallback; the diagnostic-dump path still uses the CPU combine). Honour a fixed (forced) mosaicity: SmoothMosaicityAndPartiality now overrides every frame with GetForcedMosaicity() when set, instead of always reading the per-frame integration value - so the caller can route the --mosaicity case through RotationScaleMerge (its partiality recompute makes it a natural fit) rather than a separate path. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-03 11:41:31 +02:00
leonarski_fandClaude Opus 4.8	34b3c3c4e7	RotationScaleMerge: GPU merge + error-model reductions over the resident fulls ... Port the four fulls-walking reductions of MergeAndStats to the GPU, over the fulls group-CSR already resident from scale-fulls: the per-group inv-var mean + leverage- corrected error-model samples, the merge accumulate (inv-var sums + deterministic half-sets, error-model-corrected sigma, with outlier rejection), and R_meas + the per-shell usable count. The host keeps the parts that don't parallelise cleanly or are tiny: the I2-sort + 16-bin (a,b) median fit, the per-group reject median (a per-group median is awkward on the GPU - cheap on the host from the GPU cnt), the merged export, the shells and the gemmi completeness. Only per-group arrays (~55k) + the samples (~n_fulls, for the fit) come back - the fulls are not re-walked on the host. Device HalfForImage (splitmix64) + IceRingIndex mirror the host; the corrected-sigma uses (b*I_for_b)^2 (not b^2*I^2) to match the host rounding; the R_meas usable count requires finite d (the host counts only fulls with a valid shell, and a group's fulls share d, so the shell is assigned per group). Gated on fulls_resident (GPU combine+scale-fulls active); reject is fully supported so it runs for the default rot3d command. merge+stats ~0.49 -> ~0.37s, taking RSM on lyso to ~0.78s (was ~0.91). Validated across the battery: 15/15 deterministic crystals bit-identical to the CPU path (SG / ISa / CC1.2 / completeness / total-obs, and the exact outlier-reject count), only EP_cs_01-24 noise wobbles. The em-sort + a,b fit are the remaining host floor. Non-CUDA build unaffected (use_gpu_merge is always false there). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-03 10:38:36 +02:00
leonarski_fandClaude Opus 4.8	786edf9929	RotationScaleMerge: GPU smooth-G, corr kept resident across the pass ... Apply smooth-G's corr adjustment on the device (a small kernel: corr[i] *= ratio[frame[i]] for flagged frames, double-then-float, matching host SmoothG) so the per-image corr never leaves the GPU: it now stays resident through scaling -> smooth-G -> per-frame CC -> combine, and across the two space-group passes exactly as the old host round-trip did. The host only builds the tiny per-frame ratio (g/g_smooth via the extracted ComputeSmoothGWindow) and refreshes host partials[].corr solely for the CPU-combine path (JFJOCH_RSM_CPU_COMBINE or the diagnostic dump). This drops the post-scale GetCorr and the two SetCorr re-uploads (~3x25MB/pass) plus the 6.3M host corr-adjust loop: scale-partials ~0.21->~0.10s and the smooth+combine region shrinks, taking RSM on lyso to ~0.91s (was ~1.47s with phase-1-only, ~1.71s full-CPU) - under the 1s target for this crystal; merge+stats (~0.49s) is now the dominant chunk. Bit-identical (GPU smooth-G == host SmoothG on the resident corr); validated across the battery (15/15 deterministic crystals bit-identical to CPU across default / CPU-combine / NO_GPU, only EP_cs_01-24 noise wobbles). Non-CUDA build unaffected. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-03 09:49:31 +02:00
leonarski_fandClaude Opus 4.8	87c70195f8	RotationScaleMerge: faster group-hkl via flat group ids + skip Obs stamp ... Profiling showed the per-space-group "group hkl" step (~0.30s/2-pass on lyso) is not gemmi-bound (the ASU keying is ~6ms) but memory-bandwidth-bound: stamping the group id onto, and reading it back from, the `group` field scattered across the 56-byte Obs struct touches the whole ~350MB partials array twice per pass. Precompute the per-obs AcceptReflection finiteness once (immutable) into a flat 1-byte array, then stamp the ASU-group id from rawrun_group + that flat array into a flat group_ids vector for the GPU, and build the group CSR (a stable counting sort, now parallel) from group_ids - all sequential/flat reads. The Obs.group field is written only when a CPU stage will read it (no GPU: scaling/CC/combine otherwise use group_ids / rawrun_group, never partials.group), so the default path skips the strided Obs pass entirely. group hkl ~0.31 -> ~0.20 s/2-pass on lyso. Output is bit-identical (group_ids values and the obs-index-ordered gperm are unchanged), so the merged results are unchanged; validated across the battery (15/15 deterministic crystals bit-identical to the CPU path, only EP_cs_01-24 noise keeps its benign wobble). Non-CUDA build unaffected (need_obs_group is always true there). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-03 09:25:31 +02:00
leonarski_fandClaude Opus 4.8	7d804eb799	RotationScaleMerge: GPU post-smooth group means + per-frame CC ... The single-threaded ReduceGroupMeans over the 6.3M partials (~0.07s/2-pass) and the per-frame diagnostic CC now run on the resident partials on the GPU: after SmoothG, the smoothed corr is uploaded once (and left resident for the combine, dropping the combine's redundant re-upload), then the post-smooth group means (reusing the scaling reduce) and the per-frame Pearson CC (a new one-block-per-frame kernel) run there and only the tiny per-frame cc/cc_n come back. FinalizePerFrameScale is split into ComputePerFrameCC (host reference) + the writeback; the GPU path uses ComputePartialCC. The per-frame CC is diagnostic only (the per-image scaling table), so the tree reduction's ~ulp difference from the CPU is immaterial and it does not touch merged intensities. smooth+CC region ~0.10s GPU vs ~0.15s CPU on lyso. Validated across the battery: 15/15 deterministic crystals run-to-run deterministic and merged output bit-identical to the CPU path (only EP_cs_01-24, unindexable noise, keeps its benign error-model-b wobble). CPU fallbacks (JFJOCH_RSM_CPU_COMBINE / _NO_GPU) unchanged. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-03 08:58:26 +02:00
leonarski_fandClaude Opus 4.8	617184041f	RotationScaleMerge: enable GPU combine + scale-fulls by default ... Steps 1-2 (GPU 3D-combine + resident scale-fulls) are validated bit-parity and run-to-run deterministic against the CPU path across the rotation battery, and cut the combine+scale-fulls region from ~0.46s to ~0.32s on lyso, so make them the default when a GPU is present (consistent with phase-1 partial scaling already being default-on). JFJOCH_RSM_CPU_COMBINE forces the CPU combine/scale-fulls for A/B or debugging; JFJOCH_RSM_NO_GPU still disables the whole GPU path. The only battery crystal whose reported metrics move is EP_cs_01-24 (CC1/2 2%, unindexable noise) whose upstream integration is itself nondeterministic; its merged intensities/CC/completeness are unchanged, only the ill-conditioned error-model b. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-03 08:35:23 +02:00
leonarski_fandClaude Opus 4.8	ced85bcd9d	RotationScaleMerge: GPU scale-fulls, fulls kept resident (phase 2 step 2) ... Scale the combined fulls (Unity model) on the device so they no longer round-trip between the combine and the merge: after the GPU combine, build the fulls' per-frame and per-ASU-group CSRs on the host from just the small key arrays (f_frame/f_group) with a deterministic counting sort - no GPU stable-sort - then scale in place and download once. The four scaling kernels are reused unchanged except FitPerFrameGKernel, which gains an optional `perm` argument (null for the partials, whose arrays are already frame-contiguous; a frame-grouping permutation for the emit-ordered fulls) so the fulls are scaled without a physical reorder. The Unity model falls out of giving the fulls all-ones partiality/rlp/zeta (coeff = mean), so no other kernel changes and the committed phase-1 partial-scaling path is bit-identical (perm == null -> idx == i). Validated across the rotation battery (JFJOCH_RSM_GPU_COMBINE=1): all 15 deterministic crystals stay run-to-run deterministic and their merged output is bit-identical to the CPU path (SG/ISa/CC1.2/completeness). The lone exception is EP_cs_01-24 (CC1/2 2%, R_meas 379% - unindexable noise): merged intensities/CC/completeness match exactly, but the ill-conditioned 16-bin error-model b fit amplifies the ~1e-7 scale-fulls rounding to ISa 10.6 vs 10.8 - benign, same class as the accepted phase-1 GPU rounding. The 3 upstream-nondeterministic crystals vary as before (GPU-prediction overflow, not this). Scale-fulls drops from ~0.09s to ~0 across the two passes; combine+scale-fulls region ~0.32s GPU vs ~0.46s CPU on lyso. Still opt-in (fulls are downloaded for the host merge; the win grows once the merge/error-model also stay resident). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-03 07:51:29 +02:00
leonarski_fandClaude Opus 4.8	2c928b27cd	RotationScaleMerge: GPU 3D-combine (CUDA port, phase 2 step 1) ... Port Combine() (partials->fulls) to CUDA, mirroring process_rawrun bit-for-bit: one thread per raw-hkl run splits its usable partials into rocking events (frame gap <= 2), pools background, seeds F, runs 3 de-biased Poisson reweights and the capture-uncertainty term. Emission is deterministic - a count pass, a host exclusive prefix sum for per-run offsets, then an emit pass at those offsets - so fulls come out in raw-run-major/event order, identical to the CPU path; both pass instantiations share the same arithmetic so count == emit exactly. Dmax/Dmin/Fmax reproduce std::max/min NaN semantics (not fmax) for parity. Validated across the 18-crystal rotation battery: all 15 deterministic crystals (P1/P2/C2/H3/I23/P41212/P222/P422) match the CPU combine exactly on SG/ISa/CC1.2/ completeness and run-to-run (fulls count bit-identical); the 3 upstream-nondet crystals vary from GPU-prediction overflow, not the combine. Gated opt-in behind JFJOCH_RSM_GPU_COMBINE (default = CPU combine): combine alone is timing-neutral because the shared 1.2M SortFullsByFrame std::sort dominates and the fulls round-trip adds a copy - it only pays off once the fulls stay resident for scale-fulls + merge. Also add JFJOCH_RSM_NO_GPU master switch to force the CPU fallback (incl. phase-1 scaling) from one binary for A/B parity. SortFullsByFrame extracted from the Combine tail and shared by both paths. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-03 07:27:33 +02:00
leonarski_fandClaude Opus 4.8	dd25de461d	RotationScaleMerge: GPU partial-scaling loop (CUDA port, phase 1) ... First stage of moving the rotation scale/merge onto the GPU. The per-frame partial-scaling loop (inverse-variance group-mean reduction -> robust per-frame IRLS G -> corr update, x scaling_iter) now runs in RotationScaleMergeGPU (.cu) when a GPU is present; the CPU loops remain the fallback. The host keeps the one-time raw-hkl sort and the per-space-group gemmi ASU keying, and hands the GPU a group-ordered permutation + CSR so the per-group reduction is a DETERMINISTIC segmented reduction (one thread per group, fixed order, no atomics) - preserving the run-to-run determinism just won on the CPU path (a float atomicAdd reduction would have re-introduced jitter). Reduction is one-thread-per-group (groups average tens of obs, so a block-per-group wastes threads); the IRLS is one block per frame with a deterministic shared-memory reduction. Validated: bit-identical to the CPU path and deterministic run-to-run on lyso/cytC/Ins_H/pding (P41212 ISa 7.8 CC1/2 99.7%, etc.). The scaling kernels are ~7x faster than the CPU compute (~36 ms for 3 iters vs ~0.28 s); end-to-end scale/merge ~2.0 -> ~1.5 s. The remaining gap to the <1 s target is the per-pass host round-trip (corr down/upload for the CPU combine + per-SG group-CSR rebuild); phase 2 keeps the data resident by moving the 3D combine and the merge/error-model onto the GPU too, so nothing round-trips. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-02 22:26:29 +02:00
leonarski_fandClaude Opus 4.8	29c8ba6112	rotation: deterministic frame-order mosaicity smoothing + partiality recompute ... Prediction applied a mosaicity/profile-radius moving average (RotationParameters) over the last N *processed* frames. Under the parallel per-image loop that window is thread-arrival order, so the smoothed value - and hence which reflections are predicted/integrated - was non-deterministic run-to-run, swinging CC1/2 (and even the space group) on marginal crystals. `-N 1` was deterministic; `-N 32` was not. Fix (as designed with FL): prediction now uses each frame's OWN mosaicity/profile-radius (image-local, deterministic membership - a reflection on the cutoff contributes ~nothing). The smoothing that actually matters is moved into RotationScaleMerge and done in FRAME order (deterministic): per-frame mosaicity is smoothed with the same window as smooth-G, then every partial's partiality is recomputed from it BEFORE the 3D combine. This is the mosaicity analogue of smooth-G: combining a reflection's per-frame partials only tiles the rocking curve correctly (captured fractions summing toward 1) if neighbouring frames share a consistent mosaicity. Battery (18 crystals, /data/rotation_test, 2 runs each): 15/18 now bit-identical run-to-run (the good crystals unchanged - lyso P41212 ISa 7.8 CC1/2 99.7%). The 3 residual crystals (EcwtAL500, EcwtCQ066S, pding4_003 - all large/triclinic cells) still jitter ~0.002%, traced to a SEPARATE, benign cause: the GPU prediction buffer overflow (BraggPredictionRotGPU max_reflections=10000 with a racy atomicAdd/atomicSub) on dense frames - cell/space group stay stable; to be addressed in the GPU prediction/integration rework (naively raising the cap also changes prediction quality, so it is not a one-line bump). Minor label refinements from the recomputed partiality: cytC_2 P321 -> P3121 (now consistent with cytC_3), Ins_I_2/3 report the honest I23/I213 screw-axis ambiguity. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-02 21:58:10 +02:00
leonarski_fandClaude Opus 4.8	b5d9167bf4	RotationScaleMerge: chunked parallelism -> scale/merge phase ~2s ... The per-observation corr update (7.6M items) ran through a work-stealing ParallelFor that does one atomic fetch_add PER item - pure contention for trivial work (measured: update 0.60s vs reduce 0.15s / fit 0.13s in the scale-partials loop). Add ParallelChunks (one contiguous range per worker, no per-item sync) and use it for UpdateCorr, and parallelise the ASU keying (gemmi reduction per distinct raw hkl - HKLKeyGenerator is const, safe to read concurrently) and the group-stamping over disjoint raw-hkl runs. scale-partials 0.90 -> 0.28s, group-hkl 0.20 -> 0.09s, per-pass warm 0.83s, whole scale/merge phase ~3.3 -> ~2.0s. Bit-identical output (same space group, ISa, CC1/2). ParallelChunks is the CPU stand-in for a flat CUDA grid-stride kernel; ParallelFor stays for the heavy, uneven per-frame fits where the atomic amortises and work-stealing balances the load. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-02 21:31:41 +02:00
leonarski_fandClaude Opus 4.8	29c5410328	rotation scaling: dedicated allocate-once RotationScaleMerge, ~3.4x faster scale/merge ... The rot3d post-pass (scale -> smooth-G -> 3D combine -> scale-fulls -> merge -> error model -> stats) dominated offline wall clock because ScaleOnTheFly + MergeAll + CombineRotationObservations + MergeOnTheFly rebuild a std::map keyed by hkl on every scaling iteration and every merge (7-14 map rebuilds per space-group pass), each re-walking and re-keying millions of per-frame partials. RotationScaleMerge ingests the per-frame partials ONCE into flat vectors and reuses them across both space-group passes: the raw-hkl ordering is sorted a single time at ingest, so the per-pass 3D combine only re-splits events (no sort) and the ASU grouping is one gemmi reduction per distinct raw hkl (~10x fewer) rather than per observation, reusing that order. Every hot step is a flat loop (segmented reduction + per-frame robust IRLS + parallel per-run combine) that also maps directly onto CUDA kernels. CC1/2 and the per-image CC are computed once at the end, not every iteration. It is a distinct path from ScaleOnTheFly, used only for the self-scaling rotation case (Rotation partiality + combine3d, per-image G, no B refinement, no external reference, no absorption surface, no wedge/mosaicity override). Stills, B-factor refinement, reference scaling and the absorption surface stay on the classic path. Numerically equivalent to the classic path (same robust per-frame G, same 3D combine, same XDS-order scale-fulls, same global error model, same merge statistics), validated on the 18-crystal /data/rotation_test battery: 16/18 bit-identical in space group / ISa / CC1/2; the 2 differing crystals are ice-heavy / marginal ones on which the classic path is equally non-deterministic run-to-run (a pre-existing upstream integration race). Scale/merge wall time drops ~3.4x (median 9.8 -> 4.1 s), making higher --scaling-iterations cheap. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-02 21:00:12 +02:00
leonarski_fandClaude Opus 4.8	711216e116	scan_result: short "ice" field + global Bravais lattice type ... Two per-object scan_result API tweaks (C++ internal names unchanged): - Rename the per-image scan_result field ice_ring_score -> ice (it is serialized once per image, so keep it short). Only the OpenAPI property and its regenerated C++/TS clients change; the DataMessage/EndMessage/CBOR/HDF5 field stays ice_ring_score, and the ice_ring_score plot_type enum is untouched. - Add a global rotation_bravais string (crystal-system letter + centering, e.g. "tP", "cF", "hR") to scan_result, alongside rotation_unit_cell / rotation_crystal_lattice. It comes from the RotationIndexer result (LatticeSearchResult system+centering) via the receiver, formatted by a new BravaisSymbol() helper in ScanResult.h. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-02 21:00:12 +02:00
leonarski_fandClaude Opus 4.8	ddddfb6ffc	bragg_integration: GPU box + profile-fit integrator (standalone engine) ... Reimplement BraggIntegrate2D (box sum) and ProfileIntegrate2D (Kabsch profile fit) under one roof as a base + CPU + GPU engine, mirroring the AzIntEngine / ROIIntegration pattern. Reads the preprocessed int32 ImagePreprocessorBuffer (masked=INT32_MIN, saturated=INT32_MAX), the same buffer AzIntEngineGPU/ROIIntegrationGPU consume. The CUDA engine runs one block per reflection with shared-memory reductions across six kernels (reset, mask, box-sum, profile learning, profile build, Kabsch fit); the resolution shell is computed inline. The learning/fit hot path is single precision (FP64 is throttled on consumer GPUs; reproduces the double CPU path to ~1e-4). Collapsing the per-frame CUDA API calls into one reset kernel keeps launch-latency overhead low. Standalone for now: NOT wired into IndexAndRefine. See BRAGG_INTEGRATION_ENGINE.md for the design and the binding steps. BraggIntegrationEngineGPUTest checks GPU == CPU across all three modes (box/gaussian/empirical) within numeric tolerance, plus a [bragg_bench] perf sweep. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-02 20:59:45 +02:00
leonarski_fandClaude Fable 5	347228d008	jfjoch_process: azimuthal-integration CLI + default 0.01 1/A q-spacing ... Add -q/--azim-q-spacing, --azim-min-q, --azim-max-q, --azim-phi-bins (mirroring jfjoch_azint) so offline processing can set the radial binning, applied before the azint mapping is built. Set the AzimuthalIntegrationSettings default spacing to 0.01 1/A (was 0.05): the coarse default barely resolved the narrow ice rings, diluting the ice-ring score. Finer binning sharpens it a lot with no effect on processing - EP_cs_01-17 ice score 4.6->7.3 (max 11->23), clean cytC stays ~1.0, and space group / cell / ISa / completeness are unchanged (cytC, InsI3, MyoB, pding4_001 verified full-image). Documented in JFJOCH_PROCESS.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 17:24:05 +02:00
leonarski_fandClaude Fable 5	5737d2e1d0	ice score: measure against a smooth whole-profile background, not shoulders ... The score's baseline was two adjacent shoulder bins with a bin-overlap bug - the ring's edge bins were counted in both the ring and the shoulder, since GetMeanValueOfBins is inclusive. At the typical (coarse) azint binning (dq ~ 0.05 in q, wider than the 0.03 ring half-width) a shoulder is only ~1 bin, so the ratio was noisy and poorly separated. Replace it with the ring intensity over a smooth whole-profile background: a running median of the non-ice bins, interpolated under each ring. Clean crystals now sit at ~1.0 and ice separates far more cleanly on /data/rotation_test: cytC 1.06->1.03, lysoC 1.23->2.77, EP_cs_01-17 1.67->4.51 (max 11.4). A z-score / abnormality probability was tried but is uninformative here - with many photons any real ice ring is highly significant, so the useful discriminator is the ice magnitude (this ratio), noted in CPU_DATA_ANALYSIS.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 17:05:39 +02:00
leonarski_fandClaude Fable 5	65c7eb6366	docs: complete the per-image-quantity recipe with the HDF5 read-back path ... Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 16:43:02 +02:00
leonarski_fandClaude Fable 5	0399d1799d	ice score: read back from stored HDF5 (viewer) + write to the NXmx master ... The per-image ice_ring_score was written to the data files (HDF5DataFilePluginMX) but the master-file write and the stored-file read path were missing, so opening a processed HDF5 (e.g. in jfjoch_viewer) would not surface it. Mirror bkg_estimate: write /entry/MX/iceRingScore in the NXmx master (HDF5NXmx, from EndMessage), and read it back in HDF5MetadataSource into the dataset (master + data-file reads) and per image into the message. Verified the write; the read is a byte-for-byte mirror of the working bkgEstimate round-trip. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 16:42:17 +02:00
leonarski_fandClaude Fable 5	f422988cd2	ice score: per-image ice-ring indicator plumbed through all layers ... A single per-image ice_ring_score - the strongest hexagonal-ice ring band/shoulder intensity ratio from the azimuthal profile (1 = no ice) - computed in the CPU and FPGA analysis paths and the offline azint worker, then plumbed through every layer: DataMessage/EndMessage, CBOR (frame_serialize), HDF5 (/entry/MX/iceRingScore), ScanResult, receiver plots (PlotType::IceRingScore), the OpenAPI spec (plot_type + scan_result schema, with regenerated broker/gen and frontend client) and OpenAPIConvert, the reader + Qt viewer, and the React frontend plot. Documented in docs/CBOR.md, docs/HDF5.md and docs/CPU_DATA_ANALYSIS.md, with the general "add a per-image quantity" recipe added to CLAUDE.md. Verified in HDF5: lysoC (weak ice) mean 1.23, EP_cs_01-17 (heavy ice) mean 1.67 / max 2.23. This is a monitoring quantity - it does not gate scaling (which already excludes all ice rings) or merging (handled by the CC1/2 ring mask). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 16:37:12 +02:00
leonarski_fandClaude Fable 5	6ffb16f142	merge: mask ice rings whose CC1/2 collapses below their resolution neighbours ... An ice-contaminated ring decorrelates its reflections, so its merged half-set CC1/2 falls well below the Bragg trend of its resolution shoulders (e.g. EP_cs_01-17 2.25A ring CC1/2 0.12 vs shoulders 0.89). After the final merge, compute per-hexagonal-ice-ring CC1/2 in the ring band (+/- ice width in q=2pi/d) against the shoulders either side, mask any ring more than 0.05 below its shoulders (reliable shoulder CC1/2 > 0.5, >=20 reflections each) and re-merge without them. Weak/absent rings track their neighbours and stay, so clean crystals are untouched. Data-driven and robust on /data/rotation_test: 16/18 crystals mask nothing and are unchanged (every clean crystal, plus icy-but-fine EP_cs_02-10); EP_cs_01-17 masks 6 rings, CC1/2 6.9% -> 28%; CQ066/pding4_003 mask one marginal ring each. A well-scaled CC1/2 test, replacing the fragile per-image azimuthal cutoffs. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 16:06:39 +02:00
leonarski_fandClaude Fable 5	cdbd0bed5d	partiality: fix the rotation-prediction wedge, use the stored value in scaling ... BraggPredictionRot halves settings.wedge_deg to the +/- half-wedge of the partiality erf pair, but IndexAndRefine already passed GetWedge_deg()/2, so the two /2 compounded to a half-wedge of increment/4 - half the correct Kabsch value (increment/2, which ScaleOnTheFly's RotationPartiality already used). Pass the full increment so prediction's partiality matches scaling. With prediction correct, ScaleOnTheFly now uses the stored r.partiality directly (the value the reflection was integrated with) rather than recomputing the erf pair per reflection - recomputing only when scaling overrides the geometry (-w wedge refinement, --mosaicity, or a scaling wedge override). Output-neutral on the /data/rotation_test battery (SG/cell/completeness identical, ISa/CC1/2 within run noise on the stable crystals). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 14:57:16 +02:00
leonarski_fandClaude Fable 5	8c084021e6	ice rings: exclude from scaling, keep in the final merge (recover completeness) ... --detect-ice-rings previously dropped ice-ring reflections outright before scaling, which also removed them from the merge and gutted low/mid-resolution completeness on crystals that merge them fine (e.g. lysoC 80% -> the ice bands show as jagged shell completeness). Instead flag them (new Reflection.on_ice_ring, set in JFJochProcess, carried through the 3D combine) and exclude the flag only where a model is fit: the per-image scale (ScaleOnTheFly::Accept + the per-image CC), and - for the de-novo P1 pass - the space-group search and error model (MergeOnTheFly::ExcludeIceRings, on for for_search). The final in-symmetry merge and its statistics keep them, so completeness is preserved. /data/rotation_test battery vs the previous drop-from-merge behaviour: space group correct on all 18; completeness recovered broadly with CC1/2 and ISa held (cytC_2 82->99.7%, cytC_3 73->99.7%, InsI3 76->99.5%, lysoC 80->99.7%, MyoB 80->99.7%, InsH3 78->99%). Excluding ice from the P1 search merge is what keeps the space group correct: without it InsI3 flipped I23->P1 and EP_cs_01-17 P2->P1. Known limitation: on heavy-ice crystals (EP_cs_01-17) the strong ice is garbage and keeping it in the final merge collapses CC1/2 in the ice shells (91.7->6.9%). Distinguishing strong vs weak/absent rings per crystal needs data-driven, per-ring ice detection (azimuthal radial profile) - the planned next step. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 14:29:37 +02:00
leonarski_fandClaude Fable 5	1ca5a69054	ScaleOnTheFly: solve the per-image scale by IRLS instead of Ceres ... In the default rotation (rot3d) path only G is refined - B is fixed, mosaicity is pinned and the wedge is not refined - so the predicted intensity G*coeff is linear in G and the robust (Cauchy) per-image scale is a 1-D M-estimate. Solve it directly by iteratively reweighted least squares (a few closed-form weighted ratios) instead of building a Ceres problem per image. Ceres is kept for the cases that are genuinely nonlinear: refining the B-factor or the rotation wedge. Same Cauchy objective as the Ceres path, but ~4x faster at scaling and ~30% faster overall on the /data/rotation_test battery, with space group, cell, ISa, completeness and CC1/2 matching across all 18 crystals (the two that look different, EP_cs_01-17 and EcwtAL500, are run-to-run unstable for both solvers). lyso_ref scaling 25.2->4.3s, cytC_2 15.2->2.6s, battery total 468->316s. Also drop the per-image G/B regularizers (gated by GetScalingRegularize, which nothing enables) and the now-unused RegularizationResidual. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 13:55:04 +02:00
leonarski_fandClaude Fable 5	5c9175688a	jfjoch_process: let --detect-ice-rings=on|off override the dataset ... The master file's detectorSpecific/detect_ice_rings key is read back by HDF5MetadataSource onto the experiment, so a dataset collected with ice-ring detection on was reprocessed with exclusion forced on (dropping ~9% of reflections, gutting low/mid-res completeness on clean crystals) and jfjoch_process had no way to turn it off. Make --detect-ice-rings an optional<bool> taking =on|off (bare = on, back-compatible) and apply it after the dataset value is copied in, so an explicit CLI choice wins. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 13:18:09 +02:00
leonarski_fandClaude Opus 4.8	ed9f6ac9eb	integration: remove --reciprocal-profile (proven per-frame dead-end) ... An 11-crystal mosaicity-stratified re-test (/data/rotation_test, off vs on vs a de-contaminated variant, plus a per-frame dump of the fitted widths) showed the dial is net-negative and cannot work in the per-frame paradigm: - The C|q|^2 mosaicity term - the whole point - is unfittable per-frame: the fitted curvature a2 comes out ~0 (often negative) on every crystal, with zero correlation to the XDS mosaicity (0.09..0.42 deg). Strong spots sit at low q where eta^2 q^2 is invisible; the curvature only appears at high q where there are ~0 strong spots. The law degenerates to a straight line. - With a2~0 the high-res width becomes a blind 1/cos^2(2theta) extrapolation, 2-4x wider than per-shell. The per-shell path's high-res "starvation" (flat narrow fallback) is accidentally correct: weak, crowded high-res spots want a narrow aperture, not the true wide spot shape. - The over-wide profile pulls background into weak spots -> R-meas rises, CC1/2 drops in reliable high-multiplicity shells (pding4_001, pding4_003, MyoB, EcwtCQ066). A cap at the widest well-sampled per-shell width recovers the regression, confirming over-widening is the harm. No crystal reliably wins; the apparent overall-CC gains were all in noise shells (mult 2-3, CC<20%). Delete the CLI flag, the BraggIntegrationSettings::reciprocal_profile setting, and the per-frame fit block. Default (per-shell) integration is byte-identical. NEXTGEN_INTEGRATOR.md records the finding as a dead-end for posterity. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-02 11:39:09 +02:00
leonarski_fandClaude Fable 5	7e3ad84a5b	jfjoch_process/scale: default to mmCIF, auto-detect rotation, rot3d partiality ... - Default reflection format is now mmCIF (ScalingSettings.h, the single source of truth). jfjoch_process and jfjoch_scale no longer hard-code a local MTZ default; they override only when --scaling-output is given. jfjoch_viewer inherits this. - A dataset with a rotation goniometer axis is processed as rotation data (two-pass indexing) by default; add --process-as-stills to force per-frame stills. -R still tunes the first-pass image count. - rot3d is the default partiality model for rotation processing (fixed for stills) when no explicit -P is given. - Update docs/JFJOCH_PROCESS.md: new de-novo rotation/stills examples, corrected -R/-P/--scaling-output defaults, --process-as-stills, and a real Integration table. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-02 08:34:50 +02:00
leonarski_fandClaude Opus 4.8	4b393082d4	ci: drop requests dependency and use PowerShell for the Windows release upload ... The Windows viewer runner has Python but not the 'requests' package, and does not necessarily have bash. So: - rewrite gitea_upload_file.py to use only the Python stdlib (urllib), which works with a bare interpreter on both the Linux package runners and Windows; also drop the file's unused create_release() (gitea_create_release.py owns that); - run the Windows 'Upload installer to release' step in PowerShell (always present) instead of bash, globbing the NSIS .exe with Get-ChildItem. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 21:51:54 +02:00
leonarski_fandClaude Opus 4.8	ae1ea41151	ci: matrix the Windows build and upload installers to the release on tag ... The two Windows jobs (build-windows / build-windows-nocuda) differed only in -DJFJOCH_USE_CUDA=ON/OFF, so collapse them into one matrixed job (variant cuda/nocuda), mirroring the build-rpm matrix. On a tag, upload the NSIS installer (jfjoch-<version>-win64-{cuda<major>|cpu}.exe, named in CMakeLists.txt) to the release via gitea_upload_file.py, the same helper the RPM/DEB nocuda variants use to attach viewer/writer artifacts. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 21:47:01 +02:00
leonarski_fandClaude Opus 4.8	48a7a67283	docs: publish third-party notices as a documentation page ... The THIRD_PARTY_NOTICES.md manifest lived only at the repo root and was referenced from docs/SOFTWARE.md via a ../ link that escapes the Sphinx source tree, so it never rendered in the published docs and was absent from the navigation. Add docs/THIRD_PARTY_NOTICES.md to the General toctree and fix the SOFTWARE.md link. The docs page is generated from the canonical root file by update_version.sh (like the python-client docs): licenses/*.txt links point at the repository, and the project-license links point at the in-docs LICENSE / FPGA_LICENSE pages. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 21:42:22 +02:00
leonarski_f	ec904bc21d	Remove Gitlab remains	2026-07-01 21:36:36 +02:00
leonarski_f	91cae946a9	VERSION: 1.0.0-rc.156	2026-07-01 21:33:37 +02:00
leonarski_fandClaude Opus 4.8	0f327e4550	docs: add rc.156 changelog entry ... Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 21:32:51 +02:00
leonarski_fandClaude Opus 4.8	fa121f0079	jfjoch_viewer: show "Multiple lattices detected" in image statistics ... Add a row after the Indexing solution that flags images with more than one indexed lattice (indexing_lattice_count > 1). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 21:30:10 +02:00
leonarski_fandClaude Opus 4.8	00821e32ef	rotation indexer: refine candidate cells and pick the best after refinement ... De-novo two-pass indexing failed on large-cell / superstructure / modulated crystals (EcwtAL500 0%, EcwtCQ034 0%) and mis-handled a pseudo-symmetric one (EcwtCQ066 14%). The common cause: the choice of unit cell was made too early, on the raw pre-refinement spot fraction, which is an unreliable discriminator (a correct hexagonal cell indexes only ~13% of the un-refined accumulated spots, while a wrong larger cell can index more). Move the decision to after geometry refinement, where it is reliable: - FFTIndexer now OFFERS widened candidate cells instead of deciding. ReduceResults gains a `widen` flag: the standard path (9 shortest vectors) is unchanged; the widened path anchors the two short axes and lets the third range over all filtered vectors (+dedup) to reach the long axis of an elongated cell. FilterFFTResults takes the peak count as a parameter (30 standard, 60 widened). RunInternal appends widened candidates only when its standard best indexes poorly, so compact crystals are untouched. - RotationIndexer fully refines the top few candidates and keeps the one that indexes the most spots under its own refined geometry (IndexedFraction). Each refine is length-bounded (1.2x the found cell) so a free triclinic refine cannot drift onto a pseudo-translation / modulation supercell (CQ034's satellites). The earlier (primary) candidate is preferred: a later one is adopted only if it indexes clearly more and reasonably well in absolute terms, so a twin's noisy near-tie cannot displace it. Extra/twin lattices are only searched when the chosen cell is the FFT primary (lattice[0]), since MultiLatticeSearch's rotations are derived from that primary. - The pseudo-symmetry guard (de-novo only - a user-fixed space group is always honored) is a ratio of refined fractions: refine the primitive as triclinic and drop to P only if the constrained cell indexes less than half of it. A false promotion indexes badly under its constraint (CQ066 ratio ~0.1) while genuine higher symmetry, including R-centred, indexes comparably (Ins_H R3 ratio ~0.7) and is kept. Validated on the full /data/rotation_test battery: AL500 0->89% (C2), CQ034 0->99%, CQ066 14->93% (ISa 7.2->13.7); the other 15 crystals keep their exact cell, space group, indexing rate and ISa (no regressions). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 21:17:53 +02:00
leonarski_fandClaude Opus 4.8	dfc6703f4a	jfjoch_process: don't double-smooth per-frame G on the reference path ... SmoothImageScaleG rewrites the partials in place (image_scale_corr and image_scale_g). On the no-reference path that is harmless: each scaling pass recomputes G from scratch via ScaleAllImages, so smoothing always runs on freshly-refined G. On the reference path the scaling loop is skipped, so G is computed once and stays; running scale_and_merge twice (P1 then the adopted space group) smoothed the already-smoothed G a second time, compounding into a ~2x wider effective kernel than the configured --smooth-g and biasing the merged intensities. Smooth only on the first pass of the reference path (G is unchanged afterwards, and the smoothed partials persist into the second pass's combine3D). The no-reference path is unchanged. Verified on lyso (600 frames, -P rot3d -z ref.mtz -M): the reference run now logs the smoothing once instead of twice, and the merged MTZ changes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 14:24:05 +02:00
leonarski_f	ab2176960f	OpenAPI: Fix smargon descriptions	2026-07-01 14:12:09 +02:00
leonarski_fandClaude Opus 4.8	c67ea2fa42	receiver: guard size_t underflow in compression buffer sizing ... When per-image CBOR metadata comes within 32 bytes of the buffer slot size, slot_size - (metadata_size + 32) wrapped around (size_t), passing a huge output_size to CompressImage. That defeated the buffer-too-small guard and let the compressor write the full image past the end of the slot, corrupting adjacent memory; AppendImage then threw a plain JFJochException that aborted the whole collection after the fact. Detect metadata_size + 32 >= slot_size explicitly and throw CompressionBufferTooSmallException, so the existing catch drops just this frame gracefully - the case the change was meant to handle. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 14:05:20 +02:00
leonarski_fandClaude Opus 4.8	b7613c559c	smargon: write chi/phi goniometer position into NXmx transformations ... Add an optional Smargon static positioner (chi/phi angles + rotation axes) that is reconstructed into the NXmx sample transformation chain. Chi/phi are appended at the innermost end of the chain (closest to the sample) for both the goniometer and grid-scan branches, with axes defaulting to chi {0,0,1} and phi = omega default {1,0,0}. - SmargonPosition gains chi_axis/phi_axis (common/JFJochMessages.h) - OpenAPI: optional phi_axis/chi_axis arrays; clients regenerated - OpenAPIConvert wires Dataset_settings.smargon -> DatasetSettings - CBOR serializer/deserializer round-trip the axes - tests: CBORSerialize_Start_Smargon Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 11:45:08 +02:00
leonarski_fandClaude Opus 4.8	35607057d9	CC1/2 merge: assign half-sets by deterministic per-image hash ... MergeOnTheFly::AddImage picked each image's CC1/2 half from a shared mt19937 drawn in call order (and before Mask), so the split depended on iteration/thread order and on how many images were masked. The class is mutex-guarded for concurrent "on-the-fly" use, so any parallel merge would make CC1/2 non-reproducible - a latent race. Assign the half as a splitmix64 hash of the image's stable index instead, computed after Mask. The split is now reproducible run-to-run, independent of AddImage order, parallel-safe, and decoupled from masking. Callers pass the outcome's vector index as the image id. Verified: lyso_ref two-pass -M -P rot3d gives identical CC1/2 across runs (overall 99.6%, P41212); hash split is balanced ~50/50. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-01 07:25:34 +02:00
leonarski_fandClaude Opus 4.8	a773aaa6a9	WriteReflections: write merging statistics + ISa to the mmCIF output ... The .cif (--scaling-output cif) now carries the per-shell and overall merging statistics in the standard _reflns / _reflns_shell categories (resolution, redundancy, completeness, <I/sigma>, R-rim/R-meas, CC1/2) plus the Diederichs asymptotic I/sigma (ISa) as a free-text _reflns.pdbx_diffrn_ISa item (no standard CIF tag exists). The MergeStatistics and the ISa string are threaded through WriteReflections to the mmCIF writer; jfjoch_process and jfjoch_scale pass them. Values match the text statistics table. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 22:29:45 +02:00
leonarski_fandClaude Opus 4.8	ff91ca5e3a	jfjoch_process: default --reject-outliers 6 for rot3d (XDS-like) ... Per-observation merge outlier rejection at 6 sigma is neutral-or-better across every test crystal (CC1/2 up or flat: lysoC 99.7->99.9, MyoB 98.9->99.6, EP0210 97.6->98.8, cytC 99.7->100.0, InsI2 99.2->99.7; R-meas slightly lower everywhere; ISa unchanged), and it is the lever for radiation-damaged / ice data (EP0117 reaches CC1/2 ~95% with it). Make it the rot3d default, like scale-fulls and smooth-g; --reject-outliers 0 disables it. Off for the non-rot3d partiality models, which were not benchmarked. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 22:04:36 +02:00
leonarski_fandClaude Opus 4.8	636ed1e180	space group: use merge reduced chi^2 (not R-meas) for the consistency check ... The R-meas-ratio over-symmetry guard (commit 7ade6d9) missed a near-perfect pseudo-2-fold: InsH2 (XDS R3) has a false 2-fold at operator CC 0.85 whose merged intensities still correlate well, so R-meas inflated only ~1.3x and R32 was accepted (ISa collapsing to 3.9). The reduced chi^2 (within-orbit scatter / sigma^2) is the right signal - false equivalents disagree by many sigma even when they correlate well. Measured chi^2 ratio (candidate / best subgroup) across the test set: true point groups <= 1.63 (a cubic merge of a noisy crystal, InsI3, is the worst real case), false >= 2.26 (InsH2) and 6.0 (InsH3); threshold 2.0 separates them. InsH2 now resolves to R3 (ISa 6.1 from the false R32's 3.9); InsH3/InsI2/InsI3/cytC/lysoC/MyoB/EP0210 unchanged and correct. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 21:54:51 +02:00
leonarski_fandClaude Opus 4.8	fea2790a41	rotation_indexer: detect hexagonal metric de-novo (fixes trigonal cytC indexing) ... The hexagonal lattice metric (two equal axes at 60/120 deg, both perpendicular to the third) is also satisfied by its ortho-hexagonal C-centred supercell, so the geometry-keyed LatticeSearch lands a metrically-hexagonal lattice on the C-orthorhombic setting. The -S path already re-expresses it (HexagonalConventional) keyed on the supplied space group; do the same de-novo, keyed on the metric of the reduced PRIMITIVE cell (rhombohedral lattices have a rhombohedral primitive cell and are unaffected). cytC (P3121) now indexes de-novo as hexagonal 83.8/83.8/88.6 gamma=120 and merges in P3121 (CC1/2 99.7%, ISa 13.1), for both test crystals, instead of the orthohexagonal C2/P1. lysoC/InsI3/InsH3 unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 21:31:20 +02:00
leonarski_fandClaude Opus 4.8	52248b8f5c	docs: design for auto-gated, data-driven ice-ring detection ... Records the validated z-score detector (on the azint radial background profile) and position/width estimation, for a future --detect-ice-rings=auto. Not yet implemented. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 21:22:31 +02:00
leonarski_fandClaude Opus 4.8	f3c51b0266	jfjoch_process: --min-image-cc is in percent, not a fraction (fix crash) ... --min-image-cc <num> is documented in percent and the per-image CC is a fraction in [0,1], but the value was passed straight to MinCCForImage (which requires [0,1]), so the documented usage (e.g. --min-image-cc 30) threw an uncaught JFJochException and aborted. Divide the percent argument by 100. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 21:09:18 +02:00
leonarski_fandClaude Opus 4.8	73d9e1de39	ice rings: widen the exclusion band to 0.03 q (measured ring FWHM) ... The hexagonal-ice ring FWHM measured on JUNGFRAU data (azimuthal radial background profile) is ~0.06 q, so the exclusion half-width should be ~0.03; the previous 0.02 under-covered the strong low-res rings. On EP0117 with --detect-ice-rings this lifts CC1/2 37->50%, and combined with --reject-outliers 3 (which down-weights the radiation-damaged late frames) reaches ~94% (XDS 98.5%). Only active with --detect-ice-rings, so default behaviour is unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 21:09:18 +02:00
leonarski_fandClaude Opus 4.8	a74b8767de	ice rings: add --detect-ice-rings and exclude ice-ring reflections from scaling ... jfjoch_process gains --detect-ice-rings, which (a) activates the existing spot-finder ice flagging (ice spots de-prioritised in indexing) and (b) drops reflections sitting on a hexagonal-ice powder ring from scaling/combine/merge/stats, via a new shared IsOnIceRing() helper over ICE_RING_RES_A using the spot-finder's q half-width. Their integrated intensity is contaminated by the strong, variable ice background, so leaving them in mis-scales the whole frame and inflates the error model. On EP0117 (ice-ring crystal): de-novo space-group determination recovers from P1 to P2 and CC1/2 improves (31->37%). Off by default; a no-op without the flag. This is the first, non-controversial step - the residual gap needs ice-aware background/integration (follow-up). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 20:40:51 +02:00
leonarski_fandClaude Opus 4.8	7ade6d9906	space group: reject over-symmetrised point groups by merge R-meas consistency ... Per-operator CC alone cannot separate a real weak symmetry operator from a false moderate one: a noisy crystal's genuine cubic 2-folds (InsI2 I23, CC ~0.51) score BELOW a pseudo-symmetric crystal's false 2-fold (InsH3, CC 0.64), so no CC threshold works. Add a self-consistency test: a candidate point group is accepted only if merging the intensities under it does not inflate R-meas past max_rmeas_ratio (1.5x) the most-consistent candidate. A false operator forces non-equivalent reflections together and blows R-meas up; a real one leaves it flat. Fixes InsH3: was over-called R32 (ISa collapsed to 2.2 from the forced merge), now correctly R3/H3 (ISa 10.2, matching XDS). InsI2 stays I23, and lysoC/InsI3/MyoB/ EP0210/cytC/lyso_ref are unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 20:32:03 +02:00
leonarski_fandClaude Opus 4.8	23d58b6202	integration: robust, well-conditioned per-frame --reciprocal-profile fit ... Fit the reciprocal tangential-width model y(q)=a0+a1*t+a2*t^2 in a centered, standardized variable t=(q-qbar)/qscale instead of the raw {1,q,q^2} monomials: the raw normal matrix went near-singular when the strong spots span a narrow q-range (small cell / sparse still), letting tiny per-frame jitter swing the curvature into a wild over-wide profile. Adds IRLS (Huber) robustness, a ridge on the curvature (sharp-crystal prior), and clamps the applied width to the fitted q-range (no extrapolation). Stays strictly per-frame (no dataset pooling), so it works online and for stills. Neutral on rotation data (cytC high-res CC1/2 win preserved 66.8 vs 65.6%). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 20:01:05 +02:00
leonarski_fandClaude Opus 4.8	0783af247c	scaling: express rot3d smooth-g as a rotation range in degrees (XDS DELPHI-like) ... --smooth-g now takes a rotation range in degrees (default 5) instead of a frame count, converted to an odd frame window from the oscillation step inside JFJochProcess. This makes the per-frame scale-G smoothing physical and independent of the frame slicing. (Note: G-smoothing is not the cure for the low-symmetry stochastic collapse - the per-image scale Cauchy loss is - but a degree-based window is the correct parameterization regardless.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-30 20:00:53 +02:00