// SPDX-FileCopyrightText: 2026 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only

#include <cstdlib>
#include "Rugnux.h"
#include "../image_analysis/structure_refinement/ModelValidation.h"
#include "../image_analysis/indexing/SpindleCuspLoss.h"
#include "../image_analysis/bragg_integration/SpotWidth.h"
#include "../image_analysis/bragg_integration/SpotFootprint.h"
#include "../image_analysis/bragg_integration/BraggStencil.h"   // BRAGG_FOOTPRINT_NSIGMA
#include "../image_analysis/hot_pixels/HotPixels.h"
#include "../image_analysis/SensorAbsorption.h"
#include "DiagnosticOutput.h"

#include <algorithm>
#include <atomic>
#include <chrono>
#include <cmath>
#include <condition_variable>
#include <deque>
#include <cstring>
#include <functional>
#include <future>
#include <mutex>
#include <numeric>
#include <map>
#include <set>
#include <sstream>
#include <thread>

#include "../reader/JFJochHDF5Reader.h"
#include "../common/JFJochMath.h"
#include "../common/ParallelFor.h"
#include "../common/Logger.h"
#include "../common/AzimuthalIntegrationMapping.h"
#include "../common/AzimuthalIntegrationProfile.h"
#include "../common/CUDAWrapper.h"
#include "../common/JFJochException.h"
#include "../common/time_utc.h"
#include "../image_analysis/MXAnalysisWithoutFPGA.h"
#include "../image_analysis/beam_stop/ShadowFinder.h"
#include "../image_analysis/IndexAndRefine.h"
#include "../image_analysis/indexing/AnalyzeIndexing.h"
#include "../image_analysis/indexing/HarmonicContamination.h"
#include "../image_analysis/geom_refinement/BeamCenterFFT.h"
#include "../image_analysis/geom_refinement/BeamCenterFromBackground.h"
#include "../image_analysis/geom_refinement/GeometryRefiner.h"
#include "../image_analysis/indexing/IndexerThreadPool.h"
#include "../image_analysis/spot_finding/ImageSpotFinderCPU.h"
#include "../image_analysis/spot_finding/SpotUtils.h"
#include "gemmi/twin.hpp"
#include "../image_analysis/azint/AzIntEngineCPU.h"
#include "../image_analysis/image_preprocessing/ImagePreprocessorCPU.h"
#include "../image_analysis/image_preprocessing/ImagePreprocessorBuffer.h"
#ifdef JFJOCH_USE_CUDA
#include "../image_analysis/image_preprocessing/ImagePreprocessorGPU.h"
#include "../image_analysis/image_preprocessing/ImagePreprocessorBufferGPU.h"
#include "../image_analysis/spot_finding/AdaptiveSpotFinderGPU.h"
#endif
#include "../image_analysis/scale_merge/Merge.h"
#include "../image_analysis/scale_merge/RfreeFlags.h"
#include "../image_analysis/scale_merge/RotationScaleMerge.h"
#include "../image_analysis/scale_merge/ResolutionCutoff.h"
#include "../image_analysis/scale_merge/ReindexAmbiguity.h"
#include "../image_analysis/scale_merge/ScalingResult.h"
#include "../image_analysis/scale_merge/SearchSpaceGroup.h"
#include "../image_analysis/geom_refinement/PostRefine.h"
#include "../image_analysis/lattice_search/LatticeSearch.h"
#include "../image_analysis/lattice_search/LePageLattice.h"
#include "../image_analysis/scale_merge/AnisotropyAnalysis.h"
#include "../image_analysis/scale_merge/FrenchWilson.h"
#include "../image_analysis/scale_merge/TwinningAnalysis.h"
#include "../image_analysis/scale_merge/TranslationalNCS.h"
#include "../image_analysis/scale_merge/HKLKey.h"
#include "../image_analysis/scale_merge/ScaleOnTheFly.h"
#include "../image_analysis/scale_merge/StillsPartialityRefine.h"
#include "../image_analysis/WriteReflections.h"
#include "../image_analysis/bragg_integration/CalcISigma.h"
#include "../common/Definitions.h"
#include "../common/CorrelationCoefficient.h"
#include <array>
#include <map>
#include "RugnuxPipeline.h"

using namespace rugnux_internal;

// A worker's engines hold about 24 bytes a detector pixel of its card - sixteen held 6.3 GB on a 16.8 Mpx
// detector, measured at 12-22 bytes a pixel on four data sets - and the workers keep to half of the card,
// so a small card gets fewer of them. How many there are changes only how fast a loop runs, never what it
// computes.
int Rugnux::GPUWorkersPerCard(int most) const {
    constexpr int64_t WORKER_BYTES_PER_PIXEL = 24;
    const auto fit = static_cast<int64_t>(get_gpu_total_memory(0) / 2)
                     / (WORKER_BYTES_PER_PIXEL * experiment_.GetPixelsNum());
    return static_cast<int>(std::clamp<int64_t>(fit, 1, most));
}

// Every worker submits its own kernels to a card, and a card runs out of room to accept them long before it
// runs out of work: measured on two GPUs, the image loop's own time falls to sixteen workers and then rises
// again, so forty-eight is slower than sixteen on a 16 Mpx set. Sixteen is where that measurement put the
// minimum, and a later one on a single card agrees - eight leaves the loop waiting on the queue rather than
// the card. The cap is per card because that is what the queue depth belongs to, and it only applies when -N
// was left at its default - an explicit -N is a deliberate instruction and is obeyed, which is what a
// previous attempt at this got wrong. A worker also holds GPU memory, so a small card gets fewer of them
// (GPUWorkersPerCard).
int Rugnux::ImageWorkers() const {
    const int gpus = get_gpu_count();
    if (!config_.nthreads_auto || gpus == 0)
        return config_.nthreads;
    return std::min(config_.nthreads, GPUWorkersPerCard(16) * gpus);
}

void Rugnux::ImagePass(PipelineLocals &p) {
    auto &logger = p.logger;
    auto &result = p.result;
    auto &observer = p.observer;
    auto &geometry_prepass = p.geometry_prepass;
    auto &full = p.full;
    auto &calibration = p.calibration;
    auto &per_image_analysis = p.per_image_analysis;
    auto &start_image = p.start_image;
    auto &end_image = p.end_image;
    auto &images_to_process = p.images_to_process;
    auto &dataset = p.dataset;
    auto &mapping = p.mapping;
    auto &plots = p.plots;
    auto &writer_queue = p.writer_queue;
    auto &indexer = p.indexer;
    auto &total_uncompressed_bytes = p.total_uncompressed_bytes;
    auto &end_msg = p.end_msg;
    auto &prepass_images = p.prepass_images;
    auto &loop_end = p.loop_end;
    // Main per-image loop, spread over N worker threads pulling from a shared counter. HDF5 reads
    // are serialized by the global hdf5_mutex; the analysis runs in parallel.
    std::atomic<int> next_ordinal = 0;
    std::atomic<int> finished_count = 0;
    // Images are written at their own ordinal, so a frame that fails leaves a HOLE rather than
    // shifting everything after it up. The extent of the written datasets is therefore the highest
    // ordinal reached, not the number that succeeded - counting successes makes the file one image
    // short for every failure, and the image it loses is the LAST one, not the one that failed.
    std::atomic<int> max_written_ordinal = -1;
    auto note_written = [&max_written_ordinal](int ordinal) {
        int prev = max_written_ordinal.load(std::memory_order_relaxed);
        while (prev < ordinal
               && !max_written_ordinal.compare_exchange_weak(prev, ordinal, std::memory_order_relaxed)) {
        }
    };

    // What the integrator did to the reflections it was handed, summed over the workers as each one
    // finishes (BraggIntegrationCounts). Reading an engine's counts synchronises its stream, so it is
    // done once per worker at the end rather than per image.
    std::atomic<uint64_t> bragg_predicted = 0, bragg_starved = 0, bragg_starved_nb = 0,
                          bragg_fallback = 0, bragg_overlapped = 0, bragg_overlapped_major = 0,
                          bragg_overloads_kept = 0;

    // Calibration by spots: the pooled spot list of the whole run. Ring clustering is O(n^2) in it and
    // the Hough circle centre O(n^3) in its first few hundred, so each image contributes a fair share of
    // a fixed budget instead of everything it found - a powder ring is over-determined either way.
    constexpr size_t CALIBRATION_MAX_SPOTS = 20000;
    const size_t calibration_spots_per_image =
            std::max<size_t>(1, CALIBRATION_MAX_SPOTS / static_cast<size_t>(images_to_process));
    std::mutex calibration_spot_m;
    std::vector<SpotToSave> calibration_spot_list;

    // Higher-order beam contamination, counted on the spots the lattice did not take. Each worker
    // keeps its own tally and adds it in when it is done, so the total is a sum of counts and does
    // not depend on how the images were shared out.
    std::mutex harmonic_m;
    HarmonicEvidence harmonic_evidence;

    // The crowding of the background rings at the widened radii, counted on frames spread over the
    // pass (widened_crowding_): on the geometry pre-pass, which integrates at the fixed radius while the
    // widened one is held back (bragg_adaptive_), and on a canonical pass a near tie left at the fixed
    // radius (radius_recheck_). The run made again decides the radius on the canonical pass itself.
    const std::optional<BraggIntegrationSettings> crowding_at =
            geometry_prepass ? (decide_all_late_ ? std::nullopt : bragg_adaptive_)
            : prepass_frames_only_ ? std::nullopt : radius_recheck_;
    const int crowding_stride = std::max(1, (geometry_prepass ? prepass_images : images_to_process)
                                            / CROWDING_FRAMES);
    std::mutex crowding_m;
    std::vector<std::pair<uint64_t, uint64_t>> crowding_frames;   // predicted, crowded

    // Where the indexed spots sit against their predicted reflections, measured on the geometry
    // pre-pass for the canonical pass's footprint (SpotFootprint.h).
    const bool measure_offsets = geometry_prepass && prescan_footprint_.has_value();
    std::mutex footprint_m;
    std::vector<FootprintOffset> footprint_offsets;

    auto azint_worker = [&]() {
        std::vector<uint8_t> decompression_buffer;
        std::shared_ptr<JFJochReaderRawImage> img;
        ImagePreprocessorCPU preprocessor(experiment_, pixel_mask_);
        ImagePreprocessorBuffer buffer(experiment_.GetPixelsNum());
        AzIntEngineCPU azint(mapping);
        AzimuthalIntegrationProfile profile(mapping);

        while (!cancelled_) {
            const int ordinal = next_ordinal.fetch_add(1);
            const int image_idx = start_image + ordinal * config_.stride;
            if (image_idx >= end_image) break;

            // Reuse the raw image frame to frame - a fresh 4-byte 16 Mpx frame is 72 MB of
            // untouched pages the decoder faults in one by one - except while the writer thread
            // still holds the previous one's pixels.
            if (!img || img.use_count() > 1)
                img = std::make_shared<JFJochReaderRawImage>();
            bool read = false;
            try {
                read = reader_.ReadRawImage(image_idx, *img);
            } catch (const std::exception &e) {
                if (IsFatalResourceError(e)) throw;
                logger.Error("Failed to load image {}: {}", image_idx, e.what());
                continue;
            }
            if (!read) continue;

            DataMessage msg{};
            msg.image = img->image;
            msg.number = ordinal;
            msg.original_number = image_idx;
            if (dataset->efficiency.size() > image_idx)
                msg.image_collection_efficiency = dataset->efficiency[image_idx];
            total_uncompressed_bytes += msg.image.GetUncompressedSize();

            const auto t0 = std::chrono::steady_clock::now();
            try {
                const uint8_t *image_ptr = msg.image.GetUncompressedPtr(decompression_buffer);
                preprocessor.Analyze(buffer, image_ptr, msg.image.GetMode());
                azint.Run(buffer, profile);
            } catch (const std::exception &e) {
                if (IsFatalResourceError(e)) throw;
                logger.Error("Error integrating image {}: {}", image_idx, e.what());
                continue;
            }
            msg.azint_time_s = std::chrono::duration<float>(std::chrono::steady_clock::now() - t0).count();
            msg.processing_time_s = msg.azint_time_s;
            msg.az_int_profile = profile.GetResult();
            msg.az_int_profile_count = profile.GetPixelCount();
            msg.az_int_profile_std = profile.GetStd();
            msg.bkg_estimate = profile.GetBkgEstimate(mapping.Settings());
            msg.ice_ring_score = profile.GetIceRingScore(mapping.Settings(),
                                                         config_.spot_finding.ice_ring_width_Q_recipA);
            msg.run_number = experiment_.GetRunNumber();
            msg.run_name = experiment_.GetRunName();

            plots.Add(msg, profile);
            if (writer_queue) writer_queue->Post(msg, img);
            note_written(ordinal);
            if (observer) observer->OnImageProcessed(msg);
            const int done = finished_count.fetch_add(1) + 1;
            if (observer) observer->OnProgress(done, images_to_process);
        }
    };

    // A pre-pass of the pre-pass subset (prepass_subset_only_) integrates its frames and no other.
    const std::vector<bool> subset = prepass_subset_only_ ? PrepassSubset(prepass_images) : std::vector<bool>();

    auto full_worker = [&]() {
        pin_gpu();   // round-robin per worker thread; must precede engine construction
        // Before the analysis engine, which page-locks these bytes for its uploads and unregisters
        // them when it is destroyed: it must not outlive the buffer it was given.
        std::shared_ptr<JFJochReaderRawImage> img;
        // The image the writer holds whose bytes the engine still has page-locked (see below).
        std::shared_ptr<JFJochReaderRawImage> pinned;
        MXAnalysisWithoutFPGA analysis(experiment_, mapping, pixel_mask_, *indexer,
                                       /*enable_fused_adaptive_gpu=*/true);
        AzimuthalIntegrationProfile profile(mapping);
        HarmonicEvidence worker_harmonic;
        std::vector<FootprintOffset> worker_offsets;
        std::vector<std::pair<uint64_t, uint64_t>> worker_crowding;
        if (crowding_at)
            analysis.CountCrowdingAt(*crowding_at);

        while (!cancelled_) {
            const int ordinal = next_ordinal.fetch_add(1);
            const int image_idx = start_image + ordinal * config_.stride;
            if (image_idx >= loop_end) break;
            if (!subset.empty() && !subset[ordinal])
                continue;

            // Reuse the raw image frame to frame - a fresh 4-byte 16 Mpx frame is 72 MB of
            // untouched pages the decoder faults in one by one - except while the writer thread
            // still holds the previous one's pixels.
            // The engine unregisters the old bytes when it pins the new ones (Analyze), so until then
            // the old image is kept here: the writer would otherwise free bytes still registered.
            if (!img || img.use_count() > 1) {
                pinned = std::move(img);
                img = std::make_shared<JFJochReaderRawImage>();
            }
            bool read = false;
            try {
                read = reader_.ReadRawImage(image_idx, *img);
            } catch (const std::exception &e) {
                if (IsFatalResourceError(e)) throw;
                logger.Error("Failed to load image {}: {}", image_idx, e.what());
                continue;
            }
            if (!read) continue;

            DataMessage msg{};
            msg.image = img->image;
            msg.number = ordinal;
            msg.original_number = image_idx;
            if (dataset->efficiency.size() > image_idx)
                msg.image_collection_efficiency = dataset->efficiency[image_idx];
            total_uncompressed_bytes += msg.image.GetUncompressedSize();

            const bool count_crowding = crowding_at && ordinal % crowding_stride == 0;
            analysis.CountCrowding(count_crowding);
            const BraggIntegrationCounts crowding_before = analysis.CrowdingCounts();
            const auto t0 = std::chrono::steady_clock::now();
            try {
                analysis.Analyze(msg, profile, config_.spot_finding);
            } catch (const std::exception &e) {
                if (IsFatalResourceError(e)) throw;
                logger.Error("Error analyzing image {}: {}", image_idx, e.what());
                continue;
            }
            // An indexer that failed left this frame unindexed without looking at it; the run stops
            // rather than merge what the machine dropped.
            indexer->RethrowIndexerError();
            pinned.reset();
            msg.processing_time_s = std::chrono::duration<float>(std::chrono::steady_clock::now() - t0).count();
            msg.run_number = experiment_.GetRunNumber();
            msg.run_name = experiment_.GetRunName();
            // A full analysis with indexing off is the pass of a run that found no lattice (SpotFindingPass):
            // the image was not indexed, which is said rather than left unasked.
            if (full && !config_.spot_finding.indexing)
                msg.indexing_result = false;

            if (calibration) {
                std::lock_guard lock(calibration_spot_m);
                for (size_t i = 0; i < msg.spots.size() && i < calibration_spots_per_image; ++i)
                    calibration_spot_list.push_back(msg.spots[i]);
            }

            if (count_crowding) {
                const BraggIntegrationCounts c = analysis.CrowdingCounts();
                worker_crowding.emplace_back(c.predicted - crowding_before.predicted,
                                             c.bkg_starved_by_neighbour - crowding_before.bkg_starved_by_neighbour);
            }
            AddHarmonicEvidence(msg.spots, experiment_.GetDiffractionGeometry(), worker_harmonic);
            if (measure_offsets)
                MeasureFootprintOffsets(msg.spots, msg.reflections, experiment_.GetBeamX_pxl(),
                                        experiment_.GetBeamY_pxl(), worker_offsets);

            plots.Add(msg, profile);
            if (p.scan_result) p.scan_result->Add(msg);
            if (writer_queue) writer_queue->Post(msg, img);
            note_written(ordinal);
            if (observer) observer->OnImageProcessed(msg);
            const int done = finished_count.fetch_add(1) + 1;
            if (observer) observer->OnProgress(done, images_to_process);
        }

        {
            std::lock_guard lock(harmonic_m);
            harmonic_evidence.Add(worker_harmonic);
        }
        {
            std::lock_guard lock(footprint_m);
            footprint_offsets.insert(footprint_offsets.end(), worker_offsets.begin(), worker_offsets.end());
        }
        {
            std::lock_guard lock(crowding_m);
            crowding_frames.insert(crowding_frames.end(), worker_crowding.begin(), worker_crowding.end());
        }

        const auto bragg = analysis.BraggCounts();
        bragg_predicted += bragg.predicted;
        bragg_starved += bragg.bkg_starved;
        bragg_starved_nb += bragg.bkg_starved_by_neighbour;
        bragg_fallback += bragg.profile_fallback;
        bragg_overlapped += bragg.overlapped;
        bragg_overlapped_major += bragg.overlapped_major;
        bragg_overloads_kept += bragg.overloads_kept;
    };

    if (observer) {
        observer->OnPhase("Processing images");
        observer->OnImagePass(p.write_output && !geometry_prepass);
    }
    Step(fmt::format("{}{}, {} images",
                     !per_image_analysis ? "Azimuthal integration"
                     : calibration ? "Image pass"
                     : !config_.spot_finding.indexing ? "Spot finding"
                     : experiment_.IsRotationIndexing() ? "Integration" : "Indexing and integration",
                     geometry_prepass ? " (geometry pre-pass)"
                     : full && prepass_detector_geometry_ ? " (refined geometry)" : "",
                     geometry_prepass ? prepass_images : images_to_process));

    std::function<void()> worker = per_image_analysis ? std::function<void()>(full_worker)
                                                      : std::function<void()>(azint_worker);
    // How many workers the loop actually wants (ImageWorkers). Only the full-analysis worker drives a
    // card; the azimuthal one preprocesses and integrates on the CPU and wants every thread it can have,
    // so the cap must not reach it.
    const int gpus = get_gpu_count();
    const int image_workers = per_image_analysis ? ImageWorkers() : config_.nthreads;
    if (image_workers < config_.nthreads)
        logger.Info("Per-image loop: {} of {} threads ({} GPUs) - past what a card can take queue "
                    "work from, more workers make it slower, not faster; pass -N to override",
                    image_workers, config_.nthreads, gpus);
    std::vector<std::future<void> > futures;
    futures.reserve(image_workers);
    const auto image_loop_start = std::chrono::steady_clock::now();
    for (int i = 0; i < image_workers; ++i)
        futures.push_back(std::async(std::launch::async, worker));
    for (auto &f: futures)
        f.get();
    // Everything the workers posted has to be on disk before the end message is written and the file
    // finalized, and a write that failed - out of space, above all - has to surface here rather than
    // leave a truncated file behind.
    if (writer_queue)
        writer_queue->Finish();
    // Wall time of the per-image loop alone. The per-stage means below cover only what runs inside it,
    // so without this there is nothing to compare them against and no way to see how much of a run is
    // spent outside it - on the first-pass indexing and on scaling/merging.
    result.image_loop_threads = image_workers;
    result.image_loop_time_s = std::chrono::duration<double>(
            std::chrono::steady_clock::now() - image_loop_start).count();

    result.images_processed = finished_count.load();
    // fast_live: a cut inside the canonical pass's image loop ends its sweep at the images read so far, and
    // the pass merges those, a partial sweep, as a pre-pass merges its share. Any other pass a cut stops.
    const bool partial_sweep = config_.fast_live && cancelled_ && p.write_output && !geometry_prepass
                               && result.images_processed > 0;
    result.cancelled = cancelled_ && !partial_sweep;
    if (partial_sweep)
        Note(fmt::format("cut: the sweep ends at image {} of {}", max_written_ordinal.load() + 1,
                         images_to_process));
    // Whether the pass stops here. In fast_live a cut that comes after the last image read lets the pass
    // finish - its lattice, its merge and its files are what the run has established - and only the model
    // after them is left out (ScaleMergeAndSymmetry).
    const auto stopped = [&] { return config_.fast_live ? result.cancelled : static_cast<bool>(cancelled_); };
    Note(fmt::format("{} images, {:.0f} per second", result.images_processed,
                     result.images_processed / std::max(1e-6, result.image_loop_time_s)));

    // A pre-pass that stopped at loop_end is a sweep of the frames it read, and that is all the merge and
    // the post-refinement are handed. Left in, the empty tail is a stretch of frames every per-frame
    // smoothing in the merge extrapolates over: measured, the fulls' log-scale curve, carried straight
    // across 900 empty frames, fell further every round until the scale there was zero, and the
    // cross-validation of its smoothness then took every scale in the sweep to NaN.
    if (geometry_prepass || prepass_frames_only_)
        indexer->GetIntegrationOutcome().resize(prepass_images);
    else if (partial_sweep)
        indexer->GetIntegrationOutcome().resize(max_written_ordinal.load() + 1);

    // The footprint of the reflection rather than of the spot: the pre-scan's widths plus where the
    // spots sit against their predictions (SpotFootprint.h). Like the widths it acts only where the
    // reflection outgrows the r1 disk, and like them it is held back for the canonical pass.
    if (measure_offsets && !stopped()) {
        const SpotFootprint fp = FootprintWithOffsets(*prescan_footprint_, footprint_offsets);
        const BraggIntegrationSettings canonical = bragg_adaptive_.value_or(experiment_.GetBraggIntegrationSettings());
        bool outgrows = false;
        for (size_t b = 0; b < fp.sigma_rad.size(); ++b)
            outgrows |= BRAGG_FOOTPRINT_NSIGMA * std::max(fp.sigma_rad[b], fp.sigma_tan[b]) > canonical.GetR1();
        std::string table;
        for (size_t b = 0; b < fp.sigma_rad.size(); ++b)
            table += fmt::format(" {:.0f}:{:.1f}/{:.1f}", (b + 0.5f) * fp.bin_px, fp.sigma_rad[b], fp.sigma_tan[b]);
        logger.Info("Reflection footprint: the spot widths plus the offsets of {} indexed spots from their "
                    "predicted reflections, sigma radial/tangential [px] by distance from the beam [px]:{}{}",
                    footprint_offsets.size(), table,
                    outgrows ? "" : fmt::format(" - every reflection fits the r1={:.1f} disk", canonical.GetR1()));
        if (outgrows) {
            if (!bragg_before_adaptive_)
                bragg_before_adaptive_ = experiment_.GetBraggIntegrationSettings();
            BraggIntegrationSettings bis = canonical;
            bis.Footprint(fp);
            bragg_adaptive_ = bis;
        }
    }

    if (full && indexer && experiment_.IsRotationIndexing() && !stopped()) {
        const SupercellProbe probe = indexer->GetSupercellProbe();
        if (probe[0][0].n > 0 && probe[0][1].n > 0) {
            std::string table;
            for (int sh = 0; sh < 3; sh++) {
                table += sh == 0 ? "\n  20-5 A:" : sh == 1 ? "\n  5-3 A:" : "\n  20-3 A:";
                const auto sums = [&](int c) {
                    SupercellProbeClass x = sh < 2 ? probe[c][sh] : probe[c][0];
                    if (sh == 2) x += probe[c][1];
                    return x;
                };
                const SupercellProbeFit f0 = FitSupercellProbe(sums(0));
                table += fmt::format(" 000 <I/s> {:.2f} a {:.1f} b {:.1f} <p> {:.3f} n {};", f0.mean_i_over_sigma, f0.a,
                                     f0.b, f0.mean_p, sums(0).n);
                for (int c = 1; c < 8; c++) {
                    const SupercellProbeFit f = FitSupercellProbe(sums(c));
                    table += fmt::format(" {}{}{}: {:.2f} occ {:.2f}% rock {:.2f}+-{:.2f}% flat {:.2f}%;",
                                         (c >> 2) & 1, (c >> 1) & 1, c & 1, f.mean_i_over_sigma,
                                         100.0 * f.mean_i / f0.mean_i, 100.0 * f.b / f0.b, 100.0 * f.b_se / f0.b,
                                         100.0 * f.a / f0.a);
                }
            }
            logger.Info("Supercell probe (the eight parity classes of the doubled primitive cell; <I/sigma>, "
                        "mean intensity, and the fit I = a + b p against the lattice's own):{}", table);

            // The class with the most intensity over 20-3 A goes into the report, with the cell it would
            // double the lattice to (DoubledLattice), Niggli-reduced.
            const auto band = [&](int c) { SupercellProbeClass x = probe[c][0]; x += probe[c][1]; return x; };
            const SupercellProbeFit f0 = FitSupercellProbe(band(0));
            const auto prim = indexer->GetSupercellProbePrimitive();
            int best = 0;
            for (int c = 1; c < 8; c++)
                if (best == 0 || FitSupercellProbe(band(c)).mean_i > FitSupercellProbe(band(best)).mean_i)
                    best = c;
            if (prim && f0.mean_i > 0.0 && f0.b != 0.0) {
                const SupercellProbeFit f = FitSupercellProbe(band(best));
                const std::array<int, 3> p = {(best >> 2) & 1, (best >> 1) & 1, best & 1};
                ProcessResult::SupercellClass sc;
                sc.h = p[0]; sc.k = p[1]; sc.l = p[2];
                sc.occupancy_pct = 100.0 * f.mean_i / f0.mean_i;
                sc.rock_pct = 100.0 * f.b / f0.b;
                sc.rock_se_pct = 100.0 * f.b_se / f0.b;
                sc.i_over_sigma = f.mean_i_over_sigma;
                sc.doubled_cell = DoubledLattice(*prim, p).NiggliReduce().GetUnitCell();

                // The same class out to the limit, half-integer reflections clear of the lattice's only
                // (SupercellTrend). The six other classes, which a single doubling leaves empty, are the
                // crystal's own null: measured on the same frames through the same integration, they
                // carry whatever sits at a half-integer node without being a reflection, so the class
                // is read against their median - in units of the lattice's intensity (excess) and of
                // its own standard error (z), the strongest sibling's z beside it.
                const SupercellTrend trend = indexer->GetSupercellTrend();
                const auto mean = [](const SupercellProbeClass &c) { return c.sum_i / static_cast<double>(c.n); };
                const auto se = [&](const SupercellProbeClass &c) {
                    const double m = mean(c);
                    return std::sqrt(std::max(0.0, c.sum_ii / static_cast<double>(c.n) - m * m)
                                     / static_cast<double>(c.n));
                };
                std::string trend_table;
                for (int sh = 0; sh < SUPERCELL_TREND_SHELLS; sh++) {
                    const SupercellProbeClass &l = trend[0][sh], &x = trend[best][sh];
                    std::vector<int> siblings;
                    for (int c = 1; c < 8; c++)
                        if (c != best && trend[c][sh].n > 1)
                            siblings.push_back(c);
                    if (l.n == 0 || x.n < 2 || siblings.size() < 6 || l.sum_i <= 0.0)
                        continue;
                    std::vector<double> sib_mean;
                    for (const int c : siblings)
                        sib_mean.push_back(mean(trend[c][sh]));
                    std::ranges::sort(sib_mean);
                    const double null_i = (sib_mean[2] + sib_mean[3]) / 2.0;
                    ProcessResult::SupercellClass::Shell shell;
                    shell.d_low_A = SUPERCELL_TREND_D_A[sh];
                    shell.d_high_A = SUPERCELL_TREND_D_A[sh + 1];
                    shell.n = x.n;
                    shell.lattice_i_over_sigma = l.sum_i_over_sigma / static_cast<double>(l.n);
                    shell.resolved = static_cast<double>(x.n) / static_cast<double>(l.n);
                    shell.excess = (mean(x) - null_i) / mean(l);
                    shell.z = (mean(x) - null_i) / se(x);
                    shell.sibling_z = -INFINITY;
                    for (const int c : siblings)
                        shell.sibling_z = std::max(shell.sibling_z, (mean(trend[c][sh]) - null_i) / se(trend[c][sh]));
                    sc.trend.push_back(shell);
                    trend_table += fmt::format("\n  {:.1f}-{:.1f} A: lattice n {} <I/s> {:.2f}; class n {} excess {:.3f} "
                                               "z {:.1f}, other classes z up to {:.1f}", shell.d_low_A, shell.d_high_A,
                                               l.n, shell.lattice_i_over_sigma, x.n, shell.excess, shell.z,
                                               shell.sibling_z);
                }
                logger.Info("Supercell probe to the limit, class {}{}{} (its reflections clear of the lattice's by r3, "
                            "less the median of the six other classes: mean intensity against the lattice's own, "
                            "and in standard errors):{}", p[0], p[1], p[2], trend_table);
                result.supercell = sc;
            }
        }
    }

    result.harmonic = HarmonicFromEvidence(harmonic_evidence);
    result.powder = powder_;
    result.powder_excluded_from_indexing = !config_.spot_finding.measured_ring_q_recipA.empty();
    if (result.powder_excluded_from_indexing)
        result.powder_indexing_d_min_A = config_.spot_finding.high_resolution_limit;
    if (result.harmonic.detected)
        logger.Warning("Higher-order beam contamination: {:.0f}% of the {} unindexed spots this run "
                       "could not put on the lattice are on it at one third the wavelength (the same "
                       "test off a harmonic finds {:.1f}%). They sit at {:.1f} A apparent resolution",
                       result.harmonic.fraction, result.harmonic.tested,
                       result.harmonic.null_fraction, result.harmonic.mean_d_A);

    // What the stencil cost, as the integrator measured it: the reflections dropped whole for a
    // background ring left below six clean pixels, and the rings the NEIGHBOURS' predicted regions
    // would starve where the detector alone would not - counting the tails of rocking curves the ring
    // is not kept clear of, so it measures the pattern's density. The second is what a wider r1 can
    // change, and it is what the two-pass guard reads.
    if (bragg_predicted > 0) {
        const auto frac = [&](uint64_t n) {
            return 100.0 * static_cast<double>(n) / static_cast<double>(bragg_predicted);
        };
        bkg_starved_fraction_ = static_cast<double>(bragg_starved_nb) / static_cast<double>(bragg_predicted);
        logger.Info("Integration at r1={:.1f} r2={:.1f} r3={:.2f}: of {} predicted reflections, "
                    "{:.3f}% lost their background ring; the neighbouring predictions crowd {:.3f}% of the "
                    "rings; {:.3f}% profile-fit fallbacks",
                    experiment_.GetBraggIntegrationSettings().GetR1(),
                    experiment_.GetBraggIntegrationSettings().GetR2(),
                    experiment_.GetBraggIntegrationSettings().GetR3(),
                    bragg_predicted.load(), frac(bragg_starved), frac(bragg_starved_nb),
                    frac(bragg_fallback));
        // Overlapping spots: signal disks a nearer neighbour's prediction takes pixels from, as the
        // overlap treatment decided ownership. For the report (OVERLAPPED_REFLECTIONS).
        result.predicted_reflections = bragg_predicted;
        result.overlapped_reflections = bragg_overlapped;
        result.overlapped_reflections_major = bragg_overlapped_major;
        result.overloads_kept = bragg_overloads_kept;
        if (bragg_overloads_kept > 0)
            logger.Info("Overloads: {} integrated reflections ({:.3f}% of the predicted) held a saturated pixel and are "
                        "kept as a lower bound, their saturated pixels at the saturation limit",
                        bragg_overloads_kept.load(), frac(bragg_overloads_kept));
        if (experiment_.GetBraggIntegrationSettings().GetOverlap() != OverlapMode::Off)
            logger.Info("Overlapping spots: {:.2f}% of the predicted reflections share signal pixels with a "
                        "nearer neighbour, {:.2f}% keep less than {:.0f}% of their profile as their own",
                        frac(bragg_overlapped), frac(bragg_overlapped_major),
                        100.0 * experiment_.GetBraggIntegrationSettings().GetOverlapMinPeak());
    }

    // The fraction is a ratio of sums over the counted frames, and its standard error is that ratio's,
    // from the scatter of the frames about it.
    if (crowding_at) {
        widened_crowding_.reset();
        uint64_t predicted = 0, crowded = 0;
        for (const auto &[p_i, c_i] : crowding_frames) {
            predicted += p_i;
            crowded += c_i;
        }
        if (predicted > 0 && !stopped()) {
            const double f = static_cast<double>(crowded) / static_cast<double>(predicted);
            const double n = static_cast<double>(crowding_frames.size());
            double ss = 0.0;
            for (const auto &[p_i, c_i] : crowding_frames)
                ss += std::pow(static_cast<double>(c_i) - f * static_cast<double>(p_i), 2);
            const double se = n > 1 ? std::sqrt(ss * n / (n - 1.0)) / static_cast<double>(predicted) : NAN;
            widened_crowding_ = RingCrowding{f, se, crowding_frames.size()};
            logger.Info("Counted at r1={:.1f} r2={:.2f} r3={:.2f} without integrating there, on {} images: the "
                        "neighbouring predictions would crowd {:.3f}% +- {:.3f}% of the {} rings",
                        crowding_at->GetR1(), crowding_at->GetR2(), crowding_at->GetR3(), crowding_frames.size(),
                        100.0 * f, 100.0 * se, predicted);
        }
    }

    // Every image failing is a total failure, not a run that produced nothing: it used to be reported
    // only as per-image log lines while the process still exited 0 with no output file.
    if (!stopped() && images_to_process > 0 && result.images_processed == 0)
        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
                              "No image could be analyzed - see the per-image errors above");

    result.mean_processing_time = plots.GetMeanProcessingTime();
    result.indexing_rate = plots.GetIndexingRate();
    plots.GetPlotRaw(result.image_indexed, PlotType::IndexingRate, "");

    // Powder calibration: the whole run has now been accumulated - the summed (q x azimuth) profile in
    // plots, or the pooled spot list - so fit the detector geometry to it. The caller reports it and
    // writes the PONI file; nothing here is stored in the _process.h5.
    if (calibration && !stopped())
        FitCalibration(p, calibration_spot_list);

    // End message (also written to the file).
    // The EXTENT of the per-image datasets (see note_written), not the success count: those two
    // differ by exactly the number of frames that failed, and the writer sizes its virtual datasets
    // from this.
    end_msg.max_image_number = static_cast<uint64_t>(max_written_ordinal.load() + 1);
    end_msg.images_collected_count = result.images_processed;
    end_msg.images_sent_to_write_count = result.images_processed;
    end_msg.end_date = time_UTC(std::chrono::system_clock::now());
    end_msg.run_number = experiment_.GetRunNumber();
    end_msg.run_name = experiment_.GetRunName();
    end_msg.bkg_estimate = plots.GetBkgEstimate();
    end_msg.spindle_blind_fraction = plots.GetSpindleBlindFraction();
    result.spot_resolution_estimate_A = plots.GetResolutionEstimate();
    end_msg.ice_ring_score = plots.GetIceRingScoreArray();
    end_msg.ice_ring_score_mean = plots.GetIceRingScore();
    end_msg.az_int_result["dataset"] = plots.GetAzIntProfile();
    end_msg.indexing_rate = result.indexing_rate;

    if (full && config_.spot_finding.indexing && !stopped()) {
        const auto rot = indexer->FinalizeRotationIndexing();
        indexer->RethrowIndexerError();
        if (rot.has_value()) {
            last_rotation_result_ = *rot;
            end_msg.rotation_lattice = rot->lattice;
            // The refined geometry, not the nominal StartMessage values. Two things about these five
            // numbers that are not obvious from their names. The beam centre is the PONI - the foot of
            // the perpendicular from the sample - so with a non-zero tilt it is NOT where the direct
            // beam lands; that point is D*tan(rot)/pixel away and is GetDirectBeam_pxl(). And the tilt
            // is not written here only for the record: IndexAndRefine puts it on outcome.experiment,
            // which is the geometry Bragg prediction and integration then run at. It is fitted with the
            // DISTANCE FROZEN and nothing checks the covariance, so the split between PONI and tilt is
            // not reproducible between passes even though their sum - the direct beam - is. Neither
            // number is meaningful on its own: report or feed back the whole set, or none of it.
            end_msg.refined_beam_center_x = rot->geom.GetBeamX_pxl();
            end_msg.refined_beam_center_y = rot->geom.GetBeamY_pxl();
            end_msg.refined_poni_rot1 = rot->geom.GetPoniRot1_rad();
            end_msg.refined_poni_rot2 = rot->geom.GetPoniRot2_rad();
            end_msg.refined_poni_rot3 = rot->geom.GetPoniRot3_rad();
            // rot3 is not carried: a rotation about the beam is an exact null of this experiment
            // (measured, smallest eigenvalue -5e-16), so the fit never moves it and reporting it would
            // suggest it had been determined.
            result.refined_detector_tilt_deg = std::array<double, 2>{
                rot->geom.GetPoniRot1_rad() * 180.0 / PI,
                rot->geom.GetPoniRot2_rad() * 180.0 / PI};
            if (rot->tilt_walk && rot->tilt_walk->refused)
                result.refused_detector_tilt_deg = std::array<double, 2>{
                    rot->tilt_walk->tilt_rad[0] * 180.0 / PI, rot->tilt_walk->tilt_rad[1] * 180.0 / PI};
            if (rot->axis)
                end_msg.refined_rotation_axis = rot->axis->GetAxis();
            end_msg.rotation_lattice_type = LatticeMessage{
                .centering = rot->search_result.centering,
                .niggli_class = rot->search_result.niggli_class,
                .crystal_system = rot->search_result.system
            };
            result.rotation_lattice_found = true;
            result.rotation_lattice_type = end_msg.rotation_lattice_type;
            result.twin_overlap_index = rot->twin_overlap ? rot->twin_overlap->n : 0;
            // What the equality this class asserts is worth on a REFINED cell. The indexer refines the
            // same candidate a second time with nothing held (its pseudo-symmetry guard) and hands that
            // cell over whenever the constrained one won; it comes back in the primitive reduced
            // basis, and LengthEqualityDeparture puts it in the basis where a and b are the pair this
            // class equates. Measured here, decided in RunAllPasses, which is where there is a pass to
            // compare it against.
            if (rot->unconstrained && ClassImposesLengthEquality(rot->search_result.system)) {
                const double split = LengthEqualityDeparture(rot->search_result,
                                                             rot->unconstrained->lattice);
                result.lattice_length_split = split;
                logger.Info("Rotation lattice: {}-centred {} asserts a = b; the free refinement of "
                            "the same candidate departs from that by {:.2f} %",
                            rot->search_result.centering,
                            gemmi::crystal_system_str(rot->search_result.system), 100.0 * split);
            }
            if (rot->unconstrained && (rot->search_result.system == gemmi::CrystalSystem::Monoclinic
                                       || rot->search_result.system == gemmi::CrystalSystem::Orthorhombic)) {
                const double dep = AngleEqualityDeparture_deg(rot->search_result, rot->unconstrained->lattice);
                result.lattice_angle_departure_deg = dep;
                logger.Info("Rotation lattice: {}-centred {} holds its angles at 90 deg; the free refinement "
                            "of the same candidate departs from that by {:.2f} deg",
                            rot->search_result.centering,
                            gemmi::crystal_system_str(rot->search_result.system), dep);
            }
            // The consensus cell below is this lattice's conventional cell, so this is the centring
            // it is written in - the pair the volume comparisons in Run() need.
            result.consensus_centering = rot->search_result.centering;
            // Log only: what the rest of the sweep says about the decisions taken on its first part.
            if (!p.geometry_prepass && p.write_output && experiment_.IsRotationIndexing())
                FarEndCheck(p, *rot);
        }
        result.consensus_cell = indexer->GetConsensusUnitCell();
        end_msg.unit_cell = result.consensus_cell;
    }
}