// SPDX-FileCopyrightText: 2025 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only

#include "ScaleAndMerge.h"

#include <ceres/ceres.h>

#include <thread>
#include <iostream>
#include <algorithm>
#include <cmath>
#include <limits>
#include <stdexcept>
#include <tuple>
#include <unordered_map>
#include <utility>
#include <vector>

#include "../../common/ResolutionShells.h"

namespace {

struct HKLKey {
    int64_t h = 0;
    int64_t k = 0;
    int64_t l = 0;
    bool is_positive = true; // only relevant if opt.merge_friedel == false

    bool operator==(const HKLKey& o) const noexcept {
        return h == o.h && k == o.k && l == o.l && is_positive == o.is_positive;
    }
};

struct HKLKeyHash {
    size_t operator()(const HKLKey& key) const noexcept {
        auto mix = [](uint64_t x) {
            x ^= x >> 33;
            x *= 0xff51afd7ed558ccdULL;
            x ^= x >> 33;
            x *= 0xc4ceb9fe1a85ec53ULL;
            x ^= x >> 33;
            return x;
        };
        const uint64_t a = static_cast<uint64_t>(key.h);
        const uint64_t b = static_cast<uint64_t>(key.k);
        const uint64_t c = static_cast<uint64_t>(key.l);
        const uint64_t d = static_cast<uint64_t>(key.is_positive ? 1 : 0);
        return static_cast<size_t>(mix(a) ^ (mix(b) << 1) ^ (mix(c) << 2) ^ (mix(d) << 3));
    }
};

inline int RoundImageId(float image_number, double rounding_step) {
    if (!(rounding_step > 0.0))
        rounding_step = 1.0;
    const double x = static_cast<double>(image_number) / rounding_step;
    const double r = std::round(x) * rounding_step;
    return static_cast<int>(std::llround(r / rounding_step));
}

inline double SafeSigma(double s, double min_sigma) {
    if (!std::isfinite(s) || s <= 0.0)
        return min_sigma;
    return std::max(s, min_sigma);
}

inline double SafeD(double d) {
    if (!std::isfinite(d) || d <= 0.0)
        return std::numeric_limits<double>::quiet_NaN();
    return d;
}

inline int SafeToInt(int64_t x) {
    if (x < std::numeric_limits<int>::min() || x > std::numeric_limits<int>::max())
        throw std::out_of_range("HKL index out of int range for Gemmi");
    return static_cast<int>(x);
}

inline double SafeInv(double x, double fallback) {
    if (!std::isfinite(x) || x == 0.0)
        return fallback;
    return 1.0 / x;
}

inline HKLKey CanonicalizeHKLKey(const Reflection& r, const ScaleMergeOptions& opt) {
    HKLKey key{};
    key.h = r.h;
    key.k = r.k;
    key.l = r.l;
    key.is_positive = true;

    if (!opt.space_group.has_value()) {
        if (!opt.merge_friedel) {
            const HKLKey neg{-r.h, -r.k, -r.l, true};
            const bool pos = std::tie(key.h, key.k, key.l) >= std::tie(neg.h, neg.k, neg.l);
            if (!pos) {
                key.h = -key.h;
                key.k = -key.k;
                key.l = -key.l;
                key.is_positive = false;
            }
        }
        return key;
    }

    const gemmi::SpaceGroup& sg = *opt.space_group;
    const gemmi::GroupOps gops = sg.operations();
    const gemmi::ReciprocalAsu rasu(&sg);

    const gemmi::Op::Miller in{{SafeToInt(r.h), SafeToInt(r.k), SafeToInt(r.l)}};
    const auto [asu_hkl, sign_plus] = rasu.to_asu_sign(in, gops);

    key.h = asu_hkl[0];
    key.k = asu_hkl[1];
    key.l = asu_hkl[2];
    key.is_positive = opt.merge_friedel ? true : sign_plus;
    return key;
}

/// CrystFEL-like log-scaling residual
///
///   residual = w * [ ln(I_obs) - ln(G) - ln(partiality) - ln(lp) - ln(I_true) ]
///
/// Only observations with I_obs > 0 should be fed in (the caller skips the rest).
/// G and I_true are constrained to be positive via Ceres lower bounds.
struct LogIntensityResidual {
    LogIntensityResidual(const Reflection& r, double sigma_obs, double wedge_deg, bool refine_partiality)
        : log_Iobs_(std::log(std::max(static_cast<double>(r.I), 1e-30))),
          weight_(SafeInv(sigma_obs / std::max(static_cast<double>(r.I), 1e-30), 1.0)),
          delta_phi_(r.delta_phi_deg),
          log_lp_(std::log(std::max(SafeInv(static_cast<double>(r.rlp), 1.0), 1e-30))),
          half_wedge_(wedge_deg / 2.0),
          c1_(r.zeta / std::sqrt(2.0)),
          partiality_(r.partiality),
          refine_partiality_(refine_partiality) {}

    template<typename T>
    bool operator()(const T* const G,
                    const T* const mosaicity,
                    const T* const Itrue,
                    T* residual) const {
        T partiality;
        if (refine_partiality_ && mosaicity[0] != 0.0) {
            const T arg_plus  = T(delta_phi_ + half_wedge_) * T(c1_) / mosaicity[0];
            const T arg_minus = T(delta_phi_ - half_wedge_) * T(c1_) / mosaicity[0];
            partiality = (ceres::erf(arg_plus) - ceres::erf(arg_minus)) / T(2.0);
        } else {
            partiality = T(partiality_);
        }

        // Clamp partiality away from zero so log is safe
        const T min_p = T(1e-30);
        if (partiality < min_p)
            partiality = min_p;

        // ln(I_pred) = ln(G) + ln(partiality) + ln(lp) + ln(Itrue)
        const T log_Ipred = ceres::log(G[0]) + ceres::log(partiality) + T(log_lp_) + ceres::log(Itrue[0]);
        residual[0] = (log_Ipred - T(log_Iobs_)) * T(weight_);
        return true;
    }

    double log_Iobs_;
    double weight_;       // w_i ≈ I_obs / sigma_obs  (relative weight in log-space)
    double delta_phi_;
    double log_lp_;
    double half_wedge_;
    double c1_;
    double partiality_;
    bool refine_partiality_;
};

struct IntensityResidual {
    IntensityResidual(const Reflection &r, double sigma_obs, double wedge_deg, bool refine_partiality)
        : Iobs_(static_cast<double>(r.I)),
          weight_(SafeInv(sigma_obs, 1.0)),
          delta_phi_(r.delta_phi_deg),
          lp_(SafeInv(static_cast<double>(r.rlp), 1.0)),
          half_wedge_(wedge_deg / 2.0),
          c1_(r.zeta / std::sqrt(2.0)),
          partiality_(r.partiality),
          refine_partiality_(refine_partiality) {}

    template<typename T>
    bool operator()(const T *const G,
                    const T *const mosaicity,
                    const T *const Itrue,
                    T *residual) const {
        T partiality;
        if (refine_partiality_ && mosaicity[0] != 0.0) {
            const T arg_plus = T(delta_phi_ + half_wedge_) * T(c1_) / mosaicity[0];
            const T arg_minus = T(delta_phi_ - half_wedge_) * T(c1_) / mosaicity[0];
            partiality = (ceres::erf(arg_plus) - ceres::erf(arg_minus)) / T(2.0);
        } else
            partiality = T(partiality_);

        const T Ipred = G[0] * partiality * T(lp_) * Itrue[0];
        residual[0] = (Ipred - T(Iobs_)) * T(weight_);
        return true;
    }

    double Iobs_;
    double weight_;
    double delta_phi_;
    double lp_;
    double half_wedge_;
    double c1_;
    double partiality_;
    bool refine_partiality_;
};

struct ScaleRegularizationResidual {
    explicit ScaleRegularizationResidual(double sigma_k)
        : inv_sigma_(SafeInv(sigma_k, 1.0)) {}

    template <typename T>
    bool operator()(const T* const k, T* residual) const {
        residual[0] = (k[0] - T(1.0)) * T(inv_sigma_);
        return true;
    }

    double inv_sigma_;
};

    struct SmoothnessRegularizationResidual {
        explicit SmoothnessRegularizationResidual(double sigma)
            : inv_sigma_(SafeInv(sigma, 1.0)) {}
        template <typename T>
        bool operator()(const T* const k0,
            const T* const k1,
            const T* const k2,
            T* residual) const {
            residual[0] = (ceres::log(k0[0]) + ceres::log(k2[0]) - T(2.0) * ceres::log(k1[0])) * T(inv_sigma_);
            return true;
        }

        double inv_sigma_;
    };

} // namespace

ScaleMergeResult ScaleAndMergeReflectionsCeres(const std::vector<Reflection>& observations,
                                              const ScaleMergeOptions& opt) {
    ScaleMergeResult out;

    struct ObsRef {
        const Reflection* r = nullptr;
        int img_id = 0;
        int img_slot = -1;
        int hkl_slot = -1;
        double sigma = 0.0;
    };

    std::vector<ObsRef> obs;
    obs.reserve(observations.size());

    std::unordered_map<int, int> imgIdToSlot;
    imgIdToSlot.reserve(256);

    std::unordered_map<HKLKey, int, HKLKeyHash> hklToSlot;
    hklToSlot.reserve(observations.size());

    for (const auto& r : observations) {
        const double d = SafeD(r.d);
        if (!std::isfinite(d))
            continue;
        if (!std::isfinite(r.I))
            continue;

        if (!std::isfinite(r.zeta) || r.zeta <= 0.0f)
            continue;
        if (!std::isfinite(r.rlp) || r.rlp == 0.0f)
            continue;

        const double sigma = SafeSigma(static_cast<double>(r.sigma), opt.min_sigma);

        const int img_id = RoundImageId(r.image_number, opt.image_number_rounding);

        int img_slot;
        {
            auto it = imgIdToSlot.find(img_id);
            if (it == imgIdToSlot.end()) {
                img_slot = static_cast<int>(imgIdToSlot.size());
                imgIdToSlot.emplace(img_id, img_slot);
            } else {
                img_slot = it->second;
            }
        }

        int hkl_slot;
        try {
            const HKLKey key = CanonicalizeHKLKey(r, opt);
            auto it = hklToSlot.find(key);
            if (it == hklToSlot.end()) {
                hkl_slot = static_cast<int>(hklToSlot.size());
                hklToSlot.emplace(key, hkl_slot);
            } else {
                hkl_slot = it->second;
            }
        } catch (...) {
            continue;
        }

        ObsRef o;
        o.r = &r;
        o.img_id = img_id;
        o.img_slot = img_slot;
        o.hkl_slot = hkl_slot;
        o.sigma = sigma;
        obs.push_back(o);
    }

    const int nimg = static_cast<int>(imgIdToSlot.size());
    const int nhkl = static_cast<int>(hklToSlot.size());

    out.image_scale_g.assign(nimg, 1.0);
    out.image_ids.assign(nimg, 0);

    for (const auto& kv : imgIdToSlot) {
        out.image_ids[kv.second] = kv.first;
    }

    std::vector<double> g(nimg, 1.0);
    std::vector<double> Itrue(nhkl, 0.0);

    // Mosaicity: always per-image
    std::vector<double> mosaicity(nimg, opt.mosaicity_init_deg);
    for (int i = 0; i < nimg && i < opt.mosaicity_init_deg_vec.size(); ++i) {
        if (opt.mosaicity_init_deg_vec[i] > 0.0)
            mosaicity[i] = opt.mosaicity_init_deg_vec[i];
    }

    // Initialize Itrue from per-HKL median of observed intensities
    {
        std::vector<std::vector<double>> per_hkl_I(nhkl);
        for (const auto& o : obs) {
            per_hkl_I[o.hkl_slot].push_back(static_cast<double>(o.r->I));
        }
        for (int h = 0; h < nhkl; ++h) {
            auto& v = per_hkl_I[h];
            if (v.empty()) {
                Itrue[h] = std::max(opt.min_sigma, 1e-6);
                continue;
            }
            std::nth_element(v.begin(), v.begin() + static_cast<long>(v.size() / 2), v.end());
            double med = v[v.size() / 2];
            if (!std::isfinite(med) || med <= opt.min_sigma)
                med = opt.min_sigma;
            Itrue[h] = med;
        }
    }

    ceres::Problem problem;

    const bool refine_partiality = opt.wedge_deg > 0.0;

    std::vector<bool> is_valid_hkl_slot(nhkl, false);

    for (const auto& o : obs) {
        size_t mos_slot = opt.per_image_mosaicity ? o.img_slot : 0;

        if (opt.log_scaling_residual) {
            // Log residual requires positive I_obs
            if (o.r->I <= 0.0f)
                continue;

            auto* cost = new ceres::AutoDiffCostFunction<LogIntensityResidual, 1, 1, 1, 1>(
                new LogIntensityResidual(*o.r, o.sigma, opt.wedge_deg, refine_partiality));

            problem.AddResidualBlock(cost,
                                     nullptr,
                                     &g[o.img_slot],
                                     &mosaicity[mos_slot],
                                     &Itrue[o.hkl_slot]);
            is_valid_hkl_slot[o.hkl_slot] = true;
        } else {
            auto* cost = new ceres::AutoDiffCostFunction<IntensityResidual, 1, 1, 1, 1>(
                new IntensityResidual(*o.r, o.sigma, opt.wedge_deg, refine_partiality));

            problem.AddResidualBlock(cost,
                                     nullptr,
                                     &g[o.img_slot],
                                     &mosaicity[mos_slot],
                                     &Itrue[o.hkl_slot]);
            is_valid_hkl_slot[o.hkl_slot] = true;
        }
    }

    if (opt.smoothen_g) {
        for (int i = 0; i < nimg - 2; ++i) {
            auto *cost = new ceres::AutoDiffCostFunction<SmoothnessRegularizationResidual, 1, 1, 1, 1>(
                new SmoothnessRegularizationResidual(0.05));

            problem.AddResidualBlock(cost, nullptr,
                                     &g[i],
                                     &g[i + 1],
                                     &g[i + 2]);
        }
    }

    if (opt.smoothen_mos && opt.per_image_mosaicity) {
        for (int i = 0; i < nimg - 2; ++i) {
            auto *cost = new ceres::AutoDiffCostFunction<SmoothnessRegularizationResidual, 1, 1, 1, 1>(
                new SmoothnessRegularizationResidual(0.05));

            problem.AddResidualBlock(cost, nullptr,
                                     &mosaicity[i],
                                     &mosaicity[i + 1],
                                     &mosaicity[i + 2]);
        }
    }

    // Scaling factors must be always positive
    for (int i = 0; i < nimg; ++i)
        problem.SetParameterLowerBound(&g[i], 0, 1e-12);


    // For log residual, Itrue must stay positive
    if (opt.log_scaling_residual) {
        for (int h = 0; h < nhkl; ++h) {
            if (is_valid_hkl_slot[h])
                problem.SetParameterLowerBound(&Itrue[h], 0, 1e-12);
        }
    }

    // Mosaicity refinement + bounds
    if (!opt.refine_mosaicity) {
        for (int i = 0; i < (opt.per_image_mosaicity ? nimg : 1); ++i)
            problem.SetParameterBlockConstant(&mosaicity[i]);
    } else {
        for (int i = 0; i < (opt.per_image_mosaicity ? nimg : 1); ++i) {
            problem.SetParameterLowerBound(&mosaicity[i], 0, opt.mosaicity_min_deg);
            problem.SetParameterUpperBound(&mosaicity[i], 0, opt.mosaicity_max_deg);
        }
    }

    // use all available threads
    unsigned int hw = std::thread::hardware_concurrency();
    if (hw == 0)
        hw = 1; // fallback

    ceres::Solver::Options options;

    options.linear_solver_type = ceres::SPARSE_NORMAL_CHOLESKY;
    options.minimizer_progress_to_stdout = true;
    options.max_num_iterations = opt.max_num_iterations;
    options.max_solver_time_in_seconds = opt.max_solver_time_s;
    options.num_threads = static_cast<int>(hw);

    ceres::Solver::Summary summary;
    ceres::Solve(options, &problem, &summary);

    // --- Export per-image results ---
    for (int i = 0; i < nimg; ++i)
        out.image_scale_g[i] = g[i];

    out.mosaicity_deg.resize(nimg);
    for (int i = 0; i < nimg; ++i)
        out.mosaicity_deg[i] = opt.per_image_mosaicity ? mosaicity[i] : mosaicity[0];

    std::vector<HKLKey> slotToHKL(nhkl);
    for (const auto& kv : hklToSlot) {
        slotToHKL[kv.second] = kv.first;
    }

    out.merged.resize(nhkl);
    for (int h = 0; h < nhkl; ++h) {
        out.merged[h].h = slotToHKL[h].h;
        out.merged[h].k = slotToHKL[h].k;
        out.merged[h].l = slotToHKL[h].l;
        out.merged[h].I = Itrue[h];
        out.merged[h].sigma = 0.0;
        out.merged[h].d = 0.0;
    }

    // Populate d from median of observations per HKL
    {
        std::vector<std::vector<double>> per_hkl_d(nhkl);
        for (const auto& o : obs) {
            const double d_val = static_cast<double>(o.r->d);
            if (std::isfinite(d_val) && d_val > 0.0)
                per_hkl_d[o.hkl_slot].push_back(d_val);
        }
        for (int h = 0; h < nhkl; ++h) {
            auto& v = per_hkl_d[h];
            if (!v.empty()) {
                std::nth_element(v.begin(), v.begin() + static_cast<long>(v.size() / 2), v.end());
                out.merged[h].d = v[v.size() / 2];
            }
        }
    }

    std::cout << summary.FullReport() << std::endl;

   // Merge (XDS/XSCALE style: inverse-variance weighted mean)
    {
        struct HKLAccum {
            double sum_wI = 0.0;         // Σ(w_i * I_i)
            double sum_w = 0.0;          // Σ(w_i)
            double sum_wI2 = 0.0;        // Σ(w_i * I_i²) for internal variance
            int count = 0;
        };
        std::vector<HKLAccum> accum(nhkl);

        const double half_wedge = opt.wedge_deg / 2.0;

        for (const auto& o : obs) {
            const Reflection& r = *o.r;
            const double lp = SafeInv(static_cast<double>(r.rlp), 1.0);
            const double G_i = g[o.img_slot];

            // Compute partiality with refined mosaicity
            double partiality;
            size_t mos_slot = opt.per_image_mosaicity ? o.img_slot : 0;
            if (refine_partiality && mosaicity[mos_slot] > 0.0) {
                double c1 = r.zeta / std::sqrt(2.0);
                double arg_plus  = (r.delta_phi_deg + half_wedge) * c1 / mosaicity[mos_slot];
                double arg_minus = (r.delta_phi_deg - half_wedge) * c1 / mosaicity[mos_slot];
                partiality = (std::erf(arg_plus) - std::erf(arg_minus)) / 2.0;
            } else {
                partiality = r.partiality;
            }

            if (partiality > opt.min_partiality_for_merge) {
                const double correction = G_i * partiality * lp;
                if (correction <= 0.0) continue;

                // Corrected intensity and propagated sigma
                const double I_corr = static_cast<double>(r.I) / correction;
                const double sigma_corr = o.sigma / correction;

                // XDS/XSCALE style: weight = 1/σ²
                const double w = 1.0 / (sigma_corr * sigma_corr);

                auto& a = accum[o.hkl_slot];
                a.sum_wI += w * I_corr;
                a.sum_w += w;
                a.sum_wI2 += w * I_corr * I_corr;
                a.count++;
            }
        }

        // Populate merged reflections
        for (int h = 0; h < nhkl; ++h) {
            const auto& a = accum[h];

            if (a.sum_w > 0.0) {
                // XDS/XSCALE style: weighted mean
                out.merged[h].I = a.sum_wI / a.sum_w;

                // Standard error of weighted mean: σ = 1/√(Σw)
                double sigma_stat = std::sqrt(1.0 / a.sum_w);

                if (a.count >= 2) {
                    // XDS/XSCALE style: use internal consistency to estimate sigma
                    // σ² = (1/(n-1)) * Σw_i(I_i - <I>)² / Σw_i
                    const double var_internal = a.sum_wI2 - (a.sum_wI * a.sum_wI) / a.sum_w;
                    const double sigma_int = std::sqrt(var_internal / (a.sum_w * (a.count - 1)));

                    // Take the larger of statistical and internal sigma
                    // (XDS approach: if data is worse than expected, use internal estimate)
                    out.merged[h].sigma = std::max(sigma_stat, sigma_int);
                } else {
                    // Single observation: use propagated sigma
                    out.merged[h].sigma = sigma_stat;
                }
            }
            // else: I and sigma stay at initialized values (Itrue[h] and 0.0)
        }
    }

     // ---- Compute per-shell merging statistics ----
        {
            constexpr int kStatShells = 10;

            // Determine d-range from merged reflections
            float stat_d_min = std::numeric_limits<float>::max();
            float stat_d_max = 0.0f;
            for (int h = 0; h < nhkl; ++h) {
                const auto d = static_cast<float>(out.merged[h].d);
                if (std::isfinite(d) && d > 0.0f) {
                    if (d < stat_d_min) stat_d_min = d;
                    if (d > stat_d_max) stat_d_max = d;
                }
            }

            if (stat_d_min < stat_d_max && stat_d_min > 0.0f) {
                const float d_min_pad = stat_d_min * 0.999f;
                const float d_max_pad = stat_d_max * 1.001f;
                ResolutionShells stat_shells(d_min_pad, d_max_pad, kStatShells);

                const auto shell_mean_1_d2 = stat_shells.GetShellMeanOneOverResSq();
                const auto shell_min_res   = stat_shells.GetShellMinRes();

                // Assign each unique reflection to a shell
                std::vector<int32_t> hkl_shell(nhkl, -1);
                for (int h = 0; h < nhkl; ++h) {
                    const auto d = static_cast<float>(out.merged[h].d);
                    if (std::isfinite(d) && d > 0.0f) {
                        auto s = stat_shells.GetShell(d);
                        if (s.has_value())
                            hkl_shell[h] = s.value();
                    }
                }

                // Build corrected observations per HKL (same correction logic as merge)
                struct CorrObs {
                    int hkl_slot;
                    int shell;
                    double I_corr;
                };
                std::vector<CorrObs> corr_obs;
                corr_obs.reserve(obs.size());

                const double half_wedge = opt.wedge_deg / 2.0;

                for (const auto& o : obs) {
                    if (hkl_shell[o.hkl_slot] < 0)
                        continue;

                    const Reflection& r = *o.r;
                    const double lp = SafeInv(static_cast<double>(r.rlp), 1.0);
                    const double G_i = g[o.img_slot];

                    double partiality;
                    size_t mos_slot = opt.per_image_mosaicity ? o.img_slot : 0;
                    if (refine_partiality && mosaicity[mos_slot] > 0.0) {
                        double c1 = r.zeta / std::sqrt(2.0);
                        double arg_plus  = (r.delta_phi_deg + half_wedge) * c1 / mosaicity[mos_slot];
                        double arg_minus = (r.delta_phi_deg - half_wedge) * c1 / mosaicity[mos_slot];
                        partiality = (std::erf(arg_plus) - std::erf(arg_minus)) / 2.0;
                    } else {
                        partiality = r.partiality;
                    }

                    if (partiality <= opt.min_partiality_for_merge)
                        continue;
                    const double correction = G_i * partiality * lp;
                    if (correction <= 0.0)
                        continue;

                    CorrObs co;
                    co.hkl_slot = o.hkl_slot;
                    co.shell = hkl_shell[o.hkl_slot];
                    co.I_corr = static_cast<double>(r.I) / correction;
                    corr_obs.push_back(co);
                }

                // Per-HKL: collect corrected intensities for Rmeas
                struct HKLStats {
                    double sum_I = 0.0;
                    int n = 0;
                    std::vector<double> I_list;
                };
                std::vector<HKLStats> per_hkl(nhkl);

                for (const auto& co : corr_obs) {
                    auto& hs = per_hkl[co.hkl_slot];
                    hs.sum_I += co.I_corr;
                    hs.n += 1;
                    hs.I_list.push_back(co.I_corr);
                }

                // Compute per-shell statistics
                out.statistics.shells.resize(kStatShells);

                // Accumulators per shell
                struct ShellAccum {
                    int total_obs = 0;
                    std::unordered_set<int> unique_hkls;
                    double rmeas_num = 0.0;   // Σ sqrt(n/(n-1)) * Σ|I_i - <I>|
                    double rmeas_den = 0.0;   // Σ Σ I_i
                    double sum_i_over_sig = 0.0;
                    int n_merged_with_sigma = 0;
                };
                std::vector<ShellAccum> shell_acc(kStatShells);

                // Accumulate per-HKL contributions to each shell
                for (int h = 0; h < nhkl; ++h) {
                    if (hkl_shell[h] < 0)
                        continue;
                    const int s = hkl_shell[h];
                    auto& sa = shell_acc[s];
                    const auto& hs = per_hkl[h];
                    if (hs.n == 0)
                        continue;

                    sa.unique_hkls.insert(h);
                    sa.total_obs += hs.n;

                    const double mean_I = hs.sum_I / static_cast<double>(hs.n);

                    // Rmeas numerator: sqrt(n/(n-1)) * Σ_i |I_i - <I>|
                    if (hs.n >= 2) {
                        double sum_abs_dev = 0.0;
                        for (double Ii : hs.I_list)
                            sum_abs_dev += std::abs(Ii - mean_I);
                        sa.rmeas_num += std::sqrt(static_cast<double>(hs.n) / (hs.n - 1.0)) * sum_abs_dev;
                    }

                    // Rmeas denominator: Σ_i I_i (use absolute values for robustness)
                    for (double Ii : hs.I_list)
                        sa.rmeas_den += std::abs(Ii);

                    // <I/σ> from merged reflection
                    if (out.merged[h].sigma > 0.0) {
                        sa.sum_i_over_sig += out.merged[h].I / out.merged[h].sigma;
                        sa.n_merged_with_sigma += 1;
                    }
                }

                // Completeness: enumerate ASU reflections per shell if we have space group + unit cell
                std::vector<int> possible_per_shell(kStatShells, 0);
                int total_possible = 0;
                bool have_completeness = false;

                if (opt.space_group.has_value()) {
                    // Try to build a gemmi UnitCell from the merged reflections' d-spacings
                    // We need the actual unit cell. Check if any merged reflection has d > 0.
                    // Actually, we need the real unit cell parameters. We can get them from
                    // the observations' HKL + d to deduce the metric tensor, but that's complex.
                    // Instead, we compute d from the unit cell if the caller set it via space_group.
                    // For now, we count unique reflections in the ASU by brute-force enumeration.

                    // We'll try to infer cell parameters from the merged d-spacings + HKL.
                    // Simpler: use the already-available merged reflections d-spacings to get
                    // max h, k, l and enumerate. This is a practical approximation.

                    const gemmi::SpaceGroup& sg = *opt.space_group;
                    const gemmi::GroupOps gops = sg.operations();
                    const gemmi::ReciprocalAsu rasu(&sg);

                    // Find max indices from merged data
                    int max_h = 0, max_k = 0, max_l = 0;
                    for (int idx = 0; idx < nhkl; ++idx) {
                        max_h = std::max(max_h, std::abs(out.merged[idx].h));
                        max_k = std::max(max_k, std::abs(out.merged[idx].k));
                        max_l = std::max(max_l, std::abs(out.merged[idx].l));
                    }

                    // We need a metric tensor to compute d-spacing from HKL.
                    // Estimate reciprocal metric from the merged data using least-squares.
                    // For simplicity, use the unit cell from the first observation's d + HKL
                    // This is getting complex - let's use a simpler approach:
                    // Build a map from HKL key -> d for merged reflections, then for completeness
                    // just count how many ASU reflections exist at each resolution.
                    // We enumerate all HKL in the box [-max_h..max_h] x [-max_k..max_k] x [-max_l..max_l],
                    // check if they're in the ASU, compute d from the observation data.

                    // Actually the simplest robust approach: count all ASU HKLs that have d in
                    // the range of each shell. We need d(hkl) which requires the metric tensor.
                    // Since we don't have direct access to unit cell parameters here, skip completeness
                    // or let the caller fill it in.

                    // Mark completeness as unavailable for now - the caller (jfjoch_process) can
                    // fill it in if it has the unit cell.
                    have_completeness = false;
                }

                // Fill output statistics
                for (int s = 0; s < kStatShells; ++s) {
                    auto& ss = out.statistics.shells[s];
                    const auto& sa = shell_acc[s];

                    ss.mean_one_over_d2 = shell_mean_1_d2[s];
                    // Shell boundaries: shell 0 goes from d_max_pad to shell_min_res[0], etc.
                    ss.d_min = shell_min_res[s];
                    ss.d_max = (s == 0) ? d_max_pad : shell_min_res[s - 1];

                    ss.total_observations = sa.total_obs;
                    ss.unique_reflections = static_cast<int>(sa.unique_hkls.size());
                    ss.rmeas = (sa.rmeas_den > 0.0) ? (sa.rmeas_num / sa.rmeas_den) : 0.0;
                    ss.mean_i_over_sigma = (sa.n_merged_with_sigma > 0)
                        ? (sa.sum_i_over_sig / sa.n_merged_with_sigma) : 0.0;
                    ss.possible_reflections = possible_per_shell[s];
                    ss.completeness = (have_completeness && possible_per_shell[s] > 0)
                        ? static_cast<double>(sa.unique_hkls.size()) / possible_per_shell[s] : 0.0;
                }

                // Overall statistics
                {
                    auto& ov = out.statistics.overall;
                    ov.d_min = stat_d_min;
                    ov.d_max = stat_d_max;
                    ov.mean_one_over_d2 = 0.0f;

                    int total_obs_all = 0;
                    std::unordered_set<int> all_unique;
                    double rmeas_num_all = 0.0, rmeas_den_all = 0.0;
                    double sum_isig_all = 0.0;
                    int n_isig_all = 0;

                    for (int s = 0; s < kStatShells; ++s) {
                        const auto& sa = shell_acc[s];
                        total_obs_all += sa.total_obs;
                        all_unique.insert(sa.unique_hkls.begin(), sa.unique_hkls.end());
                        rmeas_num_all += sa.rmeas_num;
                        rmeas_den_all += sa.rmeas_den;
                        sum_isig_all  += sa.sum_i_over_sig;
                        n_isig_all    += sa.n_merged_with_sigma;
                    }

                    ov.total_observations = total_obs_all;
                    ov.unique_reflections = static_cast<int>(all_unique.size());
                    ov.rmeas = (rmeas_den_all > 0.0) ? (rmeas_num_all / rmeas_den_all) : 0.0;
                    ov.mean_i_over_sigma = (n_isig_all > 0) ? (sum_isig_all / n_isig_all) : 0.0;
                    ov.possible_reflections = total_possible;
                    ov.completeness = (have_completeness && total_possible > 0)
                        ? static_cast<double>(all_unique.size()) / total_possible : 0.0;
                }
            }
        }

    return out;
}