#include <cstdint>
#include <cstdlib>
#include <cstring>
#include <cmath>
#include <vector>
#include <string>
#include <fstream>
#include <iostream>
#include <algorithm>

#ifdef _WIN32
#define BL_GEO_EXPORT __declspec(dllexport)
#include <windows.h>
#include <winhttp.h>
#else
#define BL_GEO_EXPORT __attribute__((visibility("default")))
#endif

// Constants
static const double PI = 3.14159265358979323846;
static const double DEG_TO_RAD = PI / 180.0;
static const double RAD_TO_DEG = 180.0 / PI;
static const double WGS84_A = 6378137.0;              // Semi-major axis
static const double WGS84_F = 1.0 / 298.257223563;    // Flattening
static const double WGS84_E2 = 2.0 * WGS84_F - WGS84_F * WGS84_F; // First eccentricity squared
static const double WGS84_E = 0.08181919084262149;    // First eccentricity

// Blend2D C Types
typedef uint32_t BLResult;

struct BLPoint {
    double x;
    double y;
};

struct BLMatrix2D {
    double m00;
    double m01;
    double m10;
    double m11;
    double m20;
    double m21;
};

struct BLPathCore {
    uint64_t a;
    uint64_t b;
};

struct BLContextCore {
    void* impl;
};

extern "C" {
    BLResult bl_path_init(BLPathCore* self);
    BLResult bl_path_reset(BLPathCore* self);
    BLResult bl_path_destroy(BLPathCore* self);
    BLResult bl_path_move_to(BLPathCore* self, double x0, double y0);
    BLResult bl_path_line_to(BLPathCore* self, double x1, double y1);
    BLResult bl_path_close(BLPathCore* self);

    BLResult bl_context_fill_geometry_rgba32(BLContextCore* ctx, uint32_t type, const void* data, uint32_t rgba32);
    BLResult bl_context_stroke_geometry_rgba32(BLContextCore* ctx, uint32_t type, const void* data, uint32_t rgba32);
    BLResult bl_context_set_stroke_width(BLContextCore* ctx, double width);
    BLResult bl_context_set_fill_rule(BLContextCore* ctx, uint32_t fill_rule);
    BLResult bl_context_apply_transform_op(BLContextCore* ctx, uint32_t opType, const void* opData);
}

// Geospatial Viewport Transformation
struct GeoTransform {
    double minx;
    double miny;
    double maxx;
    double maxy;
    double canvas_w;
    double canvas_h;
    double offset_x;
    double offset_y;
    double scale;
    int flip_y; // 1: standard GIS coordinate system (latitude increases upwards)
};

inline void transform_pt(const GeoTransform* gt, double gx, double gy, double& sx, double& sy) {
    if (!gt || gt->scale == 0.0) {
        sx = gx;
        sy = gy;
        return;
    }

    sx = gt->offset_x + (gx - gt->minx) * gt->scale;
    if (gt->flip_y) {
        sy = gt->canvas_h - (gt->offset_y + (gy - gt->miny) * gt->scale);
    } else {
        sy = gt->offset_y + (gy - gt->miny) * gt->scale;
    }
}

// Byte-swap helpers for Big-Endian integers in Shapefile headers
inline int32_t swap_be32(int32_t val) {
    uint32_t u = (uint32_t)val;
    return (int32_t)(((u >> 24) & 0x000000ff) |
                     ((u >> 8)  & 0x0000ff00) |
                     ((u << 8)  & 0x00ff0000) |
                     ((u << 24) & 0xff000000));
}

// ============================================================================
// ESRI Shapefile C-ABI Structs
// ============================================================================

struct ShpHeader {
    int32_t file_code;      // 9994
    int32_t file_length;    // in 16-bit words
    int32_t version;        // 1000
    int32_t shape_type;     // 1=Point, 3=PolyLine, 5=Polygon, 8=MultiPoint
    double minx, miny, maxx, maxy;
    double zmin, zmax, mmin, mmax;
    int32_t record_count;
};

struct ShpParsedData {
    int32_t shape_type;
    int32_t num_records;
    int64_t total_parts;
    int64_t total_points;
    double* coords;          // [total_points * 2] (x, y)
    int32_t* parts;          // [total_parts]
    int32_t* record_parts;   // [num_records + 1] (index into parts)
    int64_t* record_points;  // [num_records + 1] (index into points)
    double* bboxes;          // [num_records * 4] (minx, miny, maxx, maxy)
};

extern "C" {

// ============================================================================
// Projection Engine Mathematical Implementations
// ============================================================================

BL_GEO_EXPORT int bl_geo_proj_3857_forward(const double* in_coords, double* out_coords, int64_t num_points) {
    if (!in_coords || !out_coords || num_points <= 0) return -1;

    for (int64_t i = 0; i < num_points; ++i) {
        double lon = in_coords[i * 2];
        double lat = in_coords[i * 2 + 1];

        if (lat > 85.0511287798) lat = 85.0511287798;
        else if (lat < -85.0511287798) lat = -85.0511287798;

        double x = lon * (WGS84_A * DEG_TO_RAD);
        double lat_rad = lat * DEG_TO_RAD;
        double y = WGS84_A * std::log(std::tan(PI * 0.25 + lat_rad * 0.5));

        out_coords[i * 2] = x;
        out_coords[i * 2 + 1] = y;
    }
    return 0;
}

BL_GEO_EXPORT int bl_geo_proj_3857_inverse(const double* in_coords, double* out_coords, int64_t num_points) {
    if (!in_coords || !out_coords || num_points <= 0) return -1;

    for (int64_t i = 0; i < num_points; ++i) {
        double x = in_coords[i * 2];
        double y = in_coords[i * 2 + 1];

        double lon = (x / WGS84_A) * RAD_TO_DEG;
        double lat = (2.0 * std::atan(std::exp(y / WGS84_A)) - PI * 0.5) * RAD_TO_DEG;

        out_coords[i * 2] = lon;
        out_coords[i * 2 + 1] = lat;
    }
    return 0;
}

static inline double lcc_m(double phi, double e) {
    double s = std::sin(phi);
    return std::cos(phi) / std::sqrt(1.0 - e * e * s * s);
}

static inline double lcc_t(double phi, double e) {
    double s = e * std::sin(phi);
    double con = std::pow((1.0 - s) / (1.0 + s), e * 0.5);
    return std::tan(PI * 0.25 - phi * 0.5) / con;
}

BL_GEO_EXPORT int bl_geo_proj_lcc_forward(
    double phi1_deg, double phi2_deg, double phi0_deg, double lam0_deg,
    double x0, double y0,
    const double* in_coords, double* out_coords, int64_t num_points
) {
    if (!in_coords || !out_coords || num_points <= 0) return -1;

    double phi1 = phi1_deg * DEG_TO_RAD;
    double phi2 = phi2_deg * DEG_TO_RAD;
    double phi0 = phi0_deg * DEG_TO_RAD;
    double lam0 = lam0_deg * DEG_TO_RAD;
    double e = WGS84_E;

    double m1 = lcc_m(phi1, e);
    double m2 = lcc_m(phi2, e);
    double t1 = lcc_t(phi1, e);
    double t2 = lcc_t(phi2, e);
    double t0 = lcc_t(phi0, e);

    double n;
    if (std::abs(phi1 - phi2) > 1e-10) {
        n = (std::log(m1) - std::log(m2)) / (std::log(t1) - std::log(t2));
    } else {
        n = std::sin(phi1);
    }

    double F = m1 / (n * std::pow(t1, n));
    double rho0 = WGS84_A * F * std::pow(t0, n);

    for (int64_t i = 0; i < num_points; ++i) {
        double lon = in_coords[i * 2] * DEG_TO_RAD;
        double lat = in_coords[i * 2 + 1] * DEG_TO_RAD;

        if (std::abs(lat) > (PI * 0.5 - 1e-6)) {
            lat = (lat > 0 ? 1 : -1) * (PI * 0.5 - 1e-6);
        }

        double t = lcc_t(lat, e);
        double rho = WGS84_A * F * std::pow(t, n);
        double theta = n * (lon - lam0);

        out_coords[i * 2] = x0 + rho * std::sin(theta);
        out_coords[i * 2 + 1] = y0 + rho0 - rho * std::cos(theta);
    }
    return 0;
}

BL_GEO_EXPORT int bl_geo_proj_albers_forward(
    double phi1_deg, double phi2_deg, double phi0_deg, double lam0_deg,
    double x0, double y0,
    const double* in_coords, double* out_coords, int64_t num_points
) {
    if (!in_coords || !out_coords || num_points <= 0) return -1;

    double phi1 = phi1_deg * DEG_TO_RAD;
    double phi2 = phi2_deg * DEG_TO_RAD;
    double phi0 = phi0_deg * DEG_TO_RAD;
    double lam0 = lam0_deg * DEG_TO_RAD;

    double n = 0.5 * (std::sin(phi1) + std::sin(phi2));
    double C = std::cos(phi1) * std::cos(phi1) + 2.0 * n * std::sin(phi1);
    double rho0 = (WGS84_A / n) * std::sqrt(std::max(0.0, C - 2.0 * n * std::sin(phi0)));

    for (int64_t i = 0; i < num_points; ++i) {
        double lon = in_coords[i * 2] * DEG_TO_RAD;
        double lat = in_coords[i * 2 + 1] * DEG_TO_RAD;

        double theta = n * (lon - lam0);
        double q = C - 2.0 * n * std::sin(lat);
        double rho = (WGS84_A / n) * std::sqrt(std::max(0.0, q));

        out_coords[i * 2] = x0 + rho * std::sin(theta);
        out_coords[i * 2 + 1] = y0 + rho0 - rho * std::cos(theta);
    }
    return 0;
}

// ============================================================================
// Viewport & Geometry Builders
// ============================================================================

BL_GEO_EXPORT void bl_geo_init_transform(
    GeoTransform* gt,
    double minx, double miny, double maxx, double maxy,
    double canvas_w, double canvas_h,
    double padding_ratio,
    int preserve_aspect_ratio,
    int flip_y
) {
    if (!gt) return;

    gt->minx = minx;
    gt->miny = miny;
    gt->maxx = maxx;
    gt->maxy = maxy;
    gt->canvas_w = canvas_w;
    gt->canvas_h = canvas_h;
    gt->flip_y = flip_y;

    double geo_w = maxx - minx;
    double geo_h = maxy - miny;

    if (geo_w <= 0.0) geo_w = 1e-6;
    if (geo_h <= 0.0) geo_h = 1e-6;

    double pad_x = canvas_w * padding_ratio;
    double pad_y = canvas_h * padding_ratio;

    double avail_w = canvas_w - 2.0 * pad_x;
    double avail_h = canvas_h - 2.0 * pad_y;

    double sx = avail_w / geo_w;
    double sy = avail_h / geo_h;

    if (preserve_aspect_ratio) {
        gt->scale = (sx < sy) ? sx : sy;
        double used_w = geo_w * gt->scale;
        double used_h = geo_h * gt->scale;
        gt->offset_x = pad_x + (avail_w - used_w) * 0.5;
        gt->offset_y = pad_y + (avail_h - used_h) * 0.5;
    } else {
        gt->scale = sx;
        gt->offset_x = pad_x;
        gt->offset_y = pad_y;
    }
}

BL_GEO_EXPORT int bl_geo_build_polygon(
    BLPathCore* path,
    const double* coords,
    const int64_t* ring_offsets,
    int64_t num_rings,
    const GeoTransform* gt
) {
    if (!path || !coords || !ring_offsets || num_rings <= 0) return -1;

    for (int64_t r = 0; r < num_rings; ++r) {
        int64_t start_idx = ring_offsets[r];
        int64_t end_idx = ring_offsets[r + 1];
        int64_t pt_count = end_idx - start_idx;

        if (pt_count < 3) continue;

        double sx, sy;
        transform_pt(gt, coords[start_idx * 2], coords[start_idx * 2 + 1], sx, sy);
        bl_path_move_to(path, sx, sy);

        for (int64_t i = start_idx + 1; i < end_idx; ++i) {
            transform_pt(gt, coords[i * 2], coords[i * 2 + 1], sx, sy);
            bl_path_line_to(path, sx, sy);
        }

        bl_path_close(path);
    }

    return 0;
}

BL_GEO_EXPORT int bl_geo_build_linestring(
    BLPathCore* path,
    const double* coords,
    int64_t num_points,
    const GeoTransform* gt
) {
    if (!path || !coords || num_points < 2) return -1;

    double sx, sy;
    transform_pt(gt, coords[0], coords[1], sx, sy);
    bl_path_move_to(path, sx, sy);

    for (int64_t i = 1; i < num_points; ++i) {
        transform_pt(gt, coords[i * 2], coords[i * 2 + 1], sx, sy);
        bl_path_line_to(path, sx, sy);
    }

    return 0;
}

BL_GEO_EXPORT int bl_geo_build_multilinestring(
    BLPathCore* path,
    const double* coords,
    const int64_t* line_offsets,
    int64_t num_lines,
    const GeoTransform* gt
) {
    if (!path || !coords || !line_offsets || num_lines <= 0) return -1;

    for (int64_t l = 0; l < num_lines; ++l) {
        int64_t start_idx = line_offsets[l];
        int64_t end_idx = line_offsets[l + 1];
        int64_t pt_count = end_idx - start_idx;

        if (pt_count < 2) continue;

        double sx, sy;
        transform_pt(gt, coords[start_idx * 2], coords[start_idx * 2 + 1], sx, sy);
        bl_path_move_to(path, sx, sy);

        for (int64_t i = start_idx + 1; i < end_idx; ++i) {
            transform_pt(gt, coords[i * 2], coords[i * 2 + 1], sx, sy);
            bl_path_line_to(path, sx, sy);
        }
    }

    return 0;
}

BL_GEO_EXPORT int bl_geo_transform_coords(
    const double* in_coords,
    double* out_coords,
    int64_t num_points,
    const GeoTransform* gt
) {
    if (!in_coords || !out_coords || num_points <= 0) return -1;

    for (int64_t i = 0; i < num_points; ++i) {
        double sx, sy;
        transform_pt(gt, in_coords[i * 2], in_coords[i * 2 + 1], sx, sy);
        out_coords[i * 2] = sx;
        out_coords[i * 2 + 1] = sy;
    }

    return 0;
}

// ============================================================================
// High-Speed Binary Shapefile (.shp) Parser (C++)
// ============================================================================

BL_GEO_EXPORT int bl_geo_read_shp_header(const char* filepath, ShpHeader* hdr) {
    if (!filepath || !hdr) return -1;

    FILE* fp = fopen(filepath, "rb");
    if (!fp) return -2;

    uint8_t buffer[100];
    if (fread(buffer, 1, 100, fp) < 100) {
        fclose(fp);
        return -3;
    }

    int32_t file_code = swap_be32(*(int32_t*)(buffer + 0));
    if (file_code != 9994) {
        fclose(fp);
        return -4;
    }

    hdr->file_code = file_code;
    hdr->file_length = swap_be32(*(int32_t*)(buffer + 24)) * 2;
    hdr->version = *(int32_t*)(buffer + 28);
    hdr->shape_type = *(int32_t*)(buffer + 32);
    hdr->minx = *(double*)(buffer + 36);
    hdr->miny = *(double*)(buffer + 44);
    hdr->maxx = *(double*)(buffer + 52);
    hdr->maxy = *(double*)(buffer + 60);
    hdr->zmin = *(double*)(buffer + 68);
    hdr->zmax = *(double*)(buffer + 76);
    hdr->mmin = *(double*)(buffer + 84);
    hdr->mmax = *(double*)(buffer + 92);

    int count = 0;
    while (true) {
        int32_t rec_hdr[2];
        if (fread(rec_hdr, 4, 2, fp) < 2) break;
        int32_t rec_len = swap_be32(rec_hdr[1]) * 2;
        if (rec_len <= 0) continue;
        fseek(fp, rec_len, SEEK_CUR);
        count++;
    }
    hdr->record_count = count;

    fclose(fp);
    return 0;
}

BL_GEO_EXPORT ShpParsedData* bl_geo_parse_shp(const char* filepath) {
    if (!filepath) return nullptr;

    FILE* fp = fopen(filepath, "rb");
    if (!fp) return nullptr;

    uint8_t hdr_buf[100];
    if (fread(hdr_buf, 1, 100, fp) < 100) {
        fclose(fp);
        return nullptr;
    }

    int32_t shape_type = *(int32_t*)(hdr_buf + 32);

    std::vector<double> all_coords;
    std::vector<int32_t> all_parts;
    std::vector<int32_t> rec_parts;
    std::vector<int64_t> rec_points;
    std::vector<double> all_bboxes;

    rec_parts.push_back(0);
    rec_points.push_back(0);

    while (true) {
        int32_t rec_hdr[2];
        if (fread(rec_hdr, 4, 2, fp) < 2) break;

        int32_t rec_num = swap_be32(rec_hdr[0]);
        int32_t rec_len = swap_be32(rec_hdr[1]) * 2;
        if (rec_len <= 0) continue;

        std::vector<uint8_t> rec_data(rec_len);
        if (fread(rec_data.data(), 1, rec_len, fp) < (size_t)rec_len) break;

        int32_t rec_shape = *(int32_t*)(rec_data.data());
        if (rec_shape == 0) {
            all_bboxes.push_back(0); all_bboxes.push_back(0); all_bboxes.push_back(0); all_bboxes.push_back(0);
            rec_parts.push_back((int32_t)all_parts.size());
            rec_points.push_back((int64_t)(all_coords.size() / 2));
            continue;
        }

        if (rec_shape == 1 || rec_shape == 11 || rec_shape == 21) {
            double px = *(double*)(rec_data.data() + 4);
            double py = *(double*)(rec_data.data() + 12);
            all_bboxes.push_back(px); all_bboxes.push_back(py); all_bboxes.push_back(px); all_bboxes.push_back(py);
            all_coords.push_back(px);
            all_coords.push_back(py);
            all_parts.push_back((int32_t)(all_coords.size() / 2 - 1));
        } else if (rec_shape == 3 || rec_shape == 5 || rec_shape == 13 || rec_shape == 15) {
            double r_minx = *(double*)(rec_data.data() + 4);
            double r_miny = *(double*)(rec_data.data() + 12);
            double r_maxx = *(double*)(rec_data.data() + 20);
            double r_maxy = *(double*)(rec_data.data() + 28);
            all_bboxes.push_back(r_minx); all_bboxes.push_back(r_miny); all_bboxes.push_back(r_maxx); all_bboxes.push_back(r_maxy);

            int32_t num_parts = *(int32_t*)(rec_data.data() + 36);
            int32_t num_points = *(int32_t*)(rec_data.data() + 40);

            const int32_t* parts_ptr = (const int32_t*)(rec_data.data() + 44);
            const double* pts_ptr = (const double*)(rec_data.data() + 44 + num_parts * 4);

            int64_t base_pt = all_coords.size() / 2;
            for (int p = 0; p < num_parts; ++p) {
                all_parts.push_back((int32_t)(base_pt + parts_ptr[p]));
            }

            for (int i = 0; i < num_points * 2; ++i) {
                all_coords.push_back(pts_ptr[i]);
            }
        }

        rec_parts.push_back((int32_t)all_parts.size());
        rec_points.push_back((int64_t)(all_coords.size() / 2));
    }

    fclose(fp);

    auto* result = new ShpParsedData();
    result->shape_type = shape_type;
    result->num_records = (int32_t)(rec_points.size() - 1);
    result->total_parts = (int64_t)all_parts.size();
    result->total_points = (int64_t)(all_coords.size() / 2);

    result->coords = (double*)malloc(all_coords.size() * sizeof(double));
    memcpy(result->coords, all_coords.data(), all_coords.size() * sizeof(double));

    result->parts = (int32_t*)malloc(all_parts.size() * sizeof(int32_t));
    memcpy(result->parts, all_parts.data(), all_parts.size() * sizeof(int32_t));

    result->record_parts = (int32_t*)malloc(rec_parts.size() * sizeof(int32_t));
    memcpy(result->record_parts, rec_parts.data(), rec_parts.size() * sizeof(int32_t));

    result->record_points = (int64_t*)malloc(rec_points.size() * sizeof(int64_t));
    memcpy(result->record_points, rec_points.data(), rec_points.size() * sizeof(int64_t));

    result->bboxes = (double*)malloc(all_bboxes.size() * sizeof(double));
    memcpy(result->bboxes, all_bboxes.data(), all_bboxes.size() * sizeof(double));

    return result;
}

BL_GEO_EXPORT void bl_geo_free_shp_data(ShpParsedData* data) {
    if (!data) return;
    if (data->coords) free(data->coords);
    if (data->parts) free(data->parts);
    if (data->record_parts) free(data->record_parts);
    if (data->record_points) free(data->record_points);
    if (data->bboxes) free(data->bboxes);
    delete data;
}

// ============================================================================
// High-Speed HTTP/HTTPS Downloader (Native WinHTTP C-API)
// ============================================================================

BL_GEO_EXPORT int bl_geo_http_download_file(const char* url_utf8, const char* dest_path_utf8) {
#ifdef _WIN32
    if (!url_utf8 || !dest_path_utf8) return -1;

    // Convert UTF-8 URL to WCHAR
    int url_wlen = MultiByteToWideChar(CP_UTF8, 0, url_utf8, -1, NULL, 0);
    if (url_wlen <= 0) return -2;
    std::vector<WCHAR> wurl(url_wlen);
    MultiByteToWideChar(CP_UTF8, 0, url_utf8, -1, wurl.data(), url_wlen);

    URL_COMPONENTS urlComp;
    ZeroMemory(&urlComp, sizeof(urlComp));
    urlComp.dwStructSize = sizeof(urlComp);
    urlComp.dwSchemeLength    = (DWORD)-1;
    urlComp.dwHostNameLength  = (DWORD)-1;
    urlComp.dwUrlPathLength   = (DWORD)-1;
    urlComp.dwExtraInfoLength = (DWORD)-1;

    if (!WinHttpCrackUrl(wurl.data(), (DWORD)wcslen(wurl.data()), 0, &urlComp)) {
        return -3;
    }

    std::wstring hostName(urlComp.lpszHostName, urlComp.dwHostNameLength);
    std::wstring urlPath(urlComp.lpszUrlPath, urlComp.dwUrlPathLength + urlComp.dwExtraInfoLength);

    HINTERNET hSession = WinHttpOpen(L"NaraMotion-GeoEngine/1.0", WINHTTP_ACCESS_TYPE_DEFAULT_PROXY, WINHTTP_NO_PROXY_NAME, WINHTTP_NO_PROXY_BYPASS, 0);
    if (!hSession) return -4;

    HINTERNET hConnect = WinHttpConnect(hSession, hostName.c_str(), urlComp.nPort, 0);
    if (!hConnect) {
        WinHttpCloseHandle(hSession);
        return -5;
    }

    DWORD dwFlags = (urlComp.nScheme == INTERNET_SCHEME_HTTPS) ? WINHTTP_FLAG_SECURE : 0;
    HINTERNET hRequest = WinHttpOpenRequest(hConnect, L"GET", urlPath.c_str(), NULL, WINHTTP_NO_REFERER, WINHTTP_DEFAULT_ACCEPT_TYPES, dwFlags);
    if (!hRequest) {
        WinHttpCloseHandle(hConnect);
        WinHttpCloseHandle(hSession);
        return -6;
    }

    // Set auto-redirect policy
    DWORD redirectPolicy = WINHTTP_OPTION_REDIRECT_POLICY_ALWAYS;
    WinHttpSetOption(hRequest, WINHTTP_OPTION_REDIRECT_POLICY, &redirectPolicy, sizeof(redirectPolicy));

    if (!WinHttpSendRequest(hRequest, WINHTTP_NO_ADDITIONAL_HEADERS, 0, WINHTTP_NO_REQUEST_DATA, 0, 0, 0) ||
        !WinHttpReceiveResponse(hRequest, NULL)) {
        WinHttpCloseHandle(hRequest);
        WinHttpCloseHandle(hConnect);
        WinHttpCloseHandle(hSession);
        return -7;
    }

    FILE* fp = fopen(dest_path_utf8, "wb");
    if (!fp) {
        WinHttpCloseHandle(hRequest);
        WinHttpCloseHandle(hConnect);
        WinHttpCloseHandle(hSession);
        return -8;
    }

    DWORD dwSize = 0;
    std::vector<uint8_t> buffer(65536);
    while (WinHttpQueryDataAvailable(hRequest, &dwSize) && dwSize > 0) {
        DWORD dwDownloaded = 0;
        DWORD toRead = (dwSize < (DWORD)buffer.size()) ? dwSize : (DWORD)buffer.size();
        if (!WinHttpReadData(hRequest, buffer.data(), toRead, &dwDownloaded) || dwDownloaded == 0) {
            break;
        }
        fwrite(buffer.data(), 1, dwDownloaded, fp);
    }

    fclose(fp);
    WinHttpCloseHandle(hRequest);
    WinHttpCloseHandle(hConnect);
    WinHttpCloseHandle(hSession);
    return 0;
#else
    return -99; // Fallback to python on non-windows
#endif
}


// ============================================================================
// Multi-Threaded WKB / GeoParquet Geometry Decoder Engine
// ============================================================================

struct WkbParsedGeometry {
    int32_t geom_type; // 1=Point, 2=LineString, 3=Polygon, 4=MultiPoint, 5=MultiLineString, 6=MultiPolygon
    int32_t num_parts;
    int64_t num_points;
    std::vector<double> coords;
    std::vector<int32_t> part_offsets;
};

static void _parse_wkb_single(const uint8_t* data, size_t size, WkbParsedGeometry& out) {
    if (!data || size < 5) return;
    uint8_t byte_order = data[0];
    bool is_little_endian = (byte_order == 1);

    auto read_u32 = [is_little_endian](const uint8_t* ptr) -> uint32_t {
        uint32_t val;
        memcpy(&val, ptr, 4);
        if (!is_little_endian) {
            val = ((val >> 24) & 0xff) | ((val >> 8) & 0xff00) | ((val << 8) & 0xff0000) | ((val << 24) & 0xff000000);
        }
        return val;
    };

    auto read_double = [is_little_endian](const uint8_t* ptr) -> double {
        if (is_little_endian) {
            double val;
            memcpy(&val, ptr, 8);
            return val;
        } else {
            uint64_t val;
            memcpy(&val, ptr, 8);
            val = ((val >> 56) & 0x00000000000000ffULL) |
                  ((val >> 40) & 0x000000000000ff00ULL) |
                  ((val >> 24) & 0x0000000000ff0000ULL) |
                  ((val >> 8)  & 0x00000000ff000000ULL) |
                  ((val << 8)  & 0x000000ff00000000ULL) |
                  ((val << 24) & 0x0000ff0000000000ULL) |
                  ((val << 40) & 0x00ff000000000000ULL) |
                  ((val << 56) & 0xff00000000000000ULL);
            double res;
            memcpy(&res, &val, 8);
            return res;
        }
    };

    uint32_t raw_type = read_u32(data + 1);
    uint32_t base_type = raw_type % 1000;
    out.geom_type = (int32_t)base_type;

    size_t offset = 5;

    if (base_type == 1) { // Point
        if (offset + 16 <= size) {
            out.coords.push_back(read_double(data + offset));
            out.coords.push_back(read_double(data + offset + 8));
            out.num_points = 1;
            out.num_parts = 1;
            out.part_offsets.push_back(0);
        }
    } else if (base_type == 2) { // LineString
        if (offset + 4 <= size) {
            uint32_t n_pts = read_u32(data + offset);
            offset += 4;
            out.num_parts = 1;
            out.part_offsets.push_back(0);
            out.num_points = n_pts;
            out.coords.reserve(n_pts * 2);
            for (uint32_t i = 0; i < n_pts && (offset + 16 <= size); ++i) {
                out.coords.push_back(read_double(data + offset));
                out.coords.push_back(read_double(data + offset + 8));
                offset += 16;
            }
        }
    } else if (base_type == 3) { // Polygon
        if (offset + 4 <= size) {
            uint32_t n_rings = read_u32(data + offset);
            offset += 4;
            out.num_parts = n_rings;
            int64_t pt_count = 0;
            for (uint32_t r = 0; r < n_rings && (offset + 4 <= size); ++r) {
                out.part_offsets.push_back((int32_t)pt_count);
                uint32_t n_pts = read_u32(data + offset);
                offset += 4;
                pt_count += n_pts;
                for (uint32_t i = 0; i < n_pts && (offset + 16 <= size); ++i) {
                    out.coords.push_back(read_double(data + offset));
                    out.coords.push_back(read_double(data + offset + 8));
                    offset += 16;
                }
            }
            out.num_points = pt_count;
        }
    } else if (base_type == 5) { // MultiLineString
        if (offset + 4 <= size) {
            uint32_t n_lines = read_u32(data + offset);
            offset += 4;
            out.num_parts = n_lines;
            int64_t pt_count = 0;
            for (uint32_t l = 0; l < n_lines && (offset + 9 <= size); ++l) {
                out.part_offsets.push_back((int32_t)pt_count);
                offset += 5; // skip endian & sub-type
                if (offset + 4 > size) break;
                uint32_t n_pts = read_u32(data + offset);
                offset += 4;
                pt_count += n_pts;
                for (uint32_t i = 0; i < n_pts && (offset + 16 <= size); ++i) {
                    out.coords.push_back(read_double(data + offset));
                    out.coords.push_back(read_double(data + offset + 8));
                    offset += 16;
                }
            }
            out.num_points = pt_count;
        }
    } else if (base_type == 6) { // MultiPolygon
        if (offset + 4 <= size) {
            uint32_t n_polys = read_u32(data + offset);
            offset += 4;
            int64_t pt_count = 0;
            int32_t part_count = 0;
            for (uint32_t p = 0; p < n_polys && (offset + 9 <= size); ++p) {
                offset += 5; // skip endian & sub-type
                if (offset + 4 > size) break;
                uint32_t n_rings = read_u32(data + offset);
                offset += 4;
                for (uint32_t r = 0; r < n_rings && (offset + 4 <= size); ++r) {
                    out.part_offsets.push_back((int32_t)pt_count);
                    part_count++;
                    uint32_t n_pts = read_u32(data + offset);
                    offset += 4;
                    pt_count += n_pts;
                    for (uint32_t i = 0; i < n_pts && (offset + 16 <= size); ++i) {
                        out.coords.push_back(read_double(data + offset));
                        out.coords.push_back(read_double(data + offset + 8));
                        offset += 16;
                    }
                }
            }
            out.num_parts = part_count;
            out.num_points = pt_count;
        }
    }
}

BL_GEO_EXPORT int nm_geo_parse_wkb_batch(
    const uint8_t* const* wkb_ptrs, const size_t* wkb_sizes, size_t num_geoms,
    double** out_coords, int64_t* out_total_points,
    int32_t** out_parts, int64_t* out_total_parts,
    int32_t** out_geom_types, int64_t** out_geom_pt_offsets
) {
    if (!wkb_ptrs || !wkb_sizes || num_geoms == 0) return -1;

    std::vector<WkbParsedGeometry> parsed(num_geoms);

    // Parse geometries in parallel chunks
    size_t num_threads = 4;
    std::vector<size_t> chunk_starts(num_threads), chunk_ends(num_threads);
    size_t chunk_size = (num_geoms + num_threads - 1) / num_threads;

    for (size_t g = 0; g < num_geoms; ++g) {
        if (wkb_ptrs[g] && wkb_sizes[g] > 0) {
            _parse_wkb_single(wkb_ptrs[g], wkb_sizes[g], parsed[g]);
        }
    }

    int64_t tot_pts = 0;
    int64_t tot_prts = 0;
    for (size_t g = 0; g < num_geoms; ++g) {
        tot_pts += parsed[g].num_points;
        tot_prts += parsed[g].num_parts;
    }

    *out_total_points = tot_pts;
    *out_total_parts = tot_prts;

    *out_coords = (double*)malloc(tot_pts * 2 * sizeof(double));
    *out_parts = (int32_t*)malloc(tot_prts * sizeof(int32_t));
    *out_geom_types = (int32_t*)malloc(num_geoms * sizeof(int32_t));
    *out_geom_pt_offsets = (int64_t*)malloc((num_geoms + 1) * sizeof(int64_t));

    double* c_ptr = *out_coords;
    int32_t* p_ptr = *out_parts;
    int64_t curr_pt_offset = 0;
    int64_t curr_part_offset = 0;

    for (size_t g = 0; g < num_geoms; ++g) {
        (*out_geom_types)[g] = parsed[g].geom_type;
        (*out_geom_pt_offsets)[g] = curr_pt_offset;

        for (size_t p = 0; p < parsed[g].part_offsets.size(); ++p) {
            p_ptr[curr_part_offset + p] = (int32_t)(curr_pt_offset + parsed[g].part_offsets[p]);
        }
        curr_part_offset += parsed[g].part_offsets.size();

        if (parsed[g].coords.size() > 0) {
            memcpy(c_ptr + curr_pt_offset * 2, parsed[g].coords.data(), parsed[g].coords.size() * sizeof(double));
        }
        curr_pt_offset += parsed[g].num_points;
    }
    (*out_geom_pt_offsets)[num_geoms] = curr_pt_offset;

    return 0;
}

BL_GEO_EXPORT void nm_geo_free_wkb_batch(double* coords, int32_t* parts, int32_t* geom_types, int64_t* geom_pt_offsets) {
    if (coords) free(coords);
    if (parts) free(parts);
    if (geom_types) free(geom_types);
    if (geom_pt_offsets) free(geom_pt_offsets);
}


// ============================================================================
// Native Fast CRS & Metric Projection Engine
// ============================================================================

// CRS IDs: 4326=WGS84, 3857=WebMercator, 2263=NY State Plane Long Island, 5070=Albers CONUS
BL_GEO_EXPORT int nm_geo_transform_batch_crs(
    double* coords, int64_t num_points,
    int crs_from, int crs_to
) {
    if (!coords || num_points <= 0) return -1;
    if (crs_from == crs_to) return 0;

    // 1. From 4326 -> 3857
    if (crs_from == 4326 && crs_to == 3857) {
        return bl_geo_proj_3857_forward(coords, coords, num_points);
    }
    // 2. From 3857 -> 4326
    if (crs_from == 3857 && crs_to == 4326) {
        return bl_geo_proj_3857_inverse(coords, coords, num_points);
    }
    // 3. From 4326 -> 2263 (NY State Plane Long Island ftUS)
    if (crs_from == 4326 && crs_to == 2263) {
        // EPSG:2263 Standard Parameters: phi1=40°40'N, phi2=41°02'N, phi0=40°10'N, lam0=-74°00'W, x0=300000m, y0=0m
        double phi1 = 40.66666666666667;
        double phi2 = 41.03333333333333;
        double phi0 = 40.16666666666667;
        double lam0 = -74.0;
        double x0 = 300000.0;
        double y0 = 0.0;
        return bl_geo_proj_lcc_forward(phi1, phi2, phi0, lam0, x0, y0, coords, coords, num_points);
    }
    // 4. From 4326 -> 5070 (Albers Equal Area Conic CONUS)
    if (crs_from == 4326 && crs_to == 5070) {
        double phi1 = 29.5;
        double phi2 = 45.5;
        double phi0 = 23.0;
        double lam0 = -96.0;
        double x0 = 0.0;
        double y0 = 0.0;
        return bl_geo_proj_albers_forward(phi1, phi2, phi0, lam0, x0, y0, coords, coords, num_points);
    }

    return -2; // Unsupported combination
}


// ============================================================================
// Hardware-Accelerated Bounding Box Liang-Barsky Vector Clipping
// ============================================================================

static inline bool _clip_line_liang_barsky(
    double minx, double miny, double maxx, double maxy,
    double x0, double y0, double x1, double y1,
    double& cx0, double& cy0, double& cx1, double& cy1
) {
    double p[4] = { -(x1 - x0), x1 - x0, -(y1 - y0), y1 - y0 };
    double q[4] = { x0 - minx, maxx - x0, y0 - miny, maxy - y0 };
    double u1 = 0.0, u2 = 1.0;

    for (int i = 0; i < 4; ++i) {
        if (p[i] == 0.0) {
            if (q[i] < 0.0) return false;
        } else {
            double t = q[i] / p[i];
            if (p[i] < 0.0) {
                if (t > u1) u1 = t;
            } else {
                if (t < u2) u2 = t;
            }
        }
    }

    if (u1 > u2) return false;

    cx0 = x0 + u1 * (x1 - x0);
    cy0 = y0 + u1 * (y1 - y0);
    cx1 = x0 + u2 * (x1 - x0);
    cy1 = y0 + u2 * (y1 - y0);
    return true;
}

BL_GEO_EXPORT int nm_geo_clip_lines_bbox(
    const double* in_coords, int64_t num_segments,
    double minx, double miny, double maxx, double maxy,
    double* out_coords, int64_t* out_count
) {
    if (!in_coords || num_segments <= 0 || !out_coords || !out_count) return -1;

    int64_t count = 0;
    for (int64_t i = 0; i < num_segments; ++i) {
        double x0 = in_coords[i * 4];
        double y0 = in_coords[i * 4 + 1];
        double x1 = in_coords[i * 4 + 2];
        double y1 = in_coords[i * 4 + 3];

        double cx0, cy0, cx1, cy1;
        if (_clip_line_liang_barsky(minx, miny, maxx, maxy, x0, y0, x1, y1, cx0, cy0, cx1, cy1)) {
            out_coords[count * 4] = cx0;
            out_coords[count * 4 + 1] = cy0;
            out_coords[count * 4 + 2] = cx1;
            out_coords[count * 4 + 3] = cy1;
            count++;
        }
    }

    *out_count = count;
    return 0;
}


// ============================================================================
// Procedural Spatial Animation & Wave Propagation Generator
// ============================================================================

BL_GEO_EXPORT int nm_geo_calculate_propagation(
    const double* centroids, size_t num_items,
    const double* epicenters, size_t num_epicenters,
    double duration_frames, double speed,
    double* out_start_frames, double* out_end_frames
) {
    if (!centroids || num_items == 0 || !epicenters || num_epicenters == 0 || !out_start_frames || !out_end_frames) {
        return -1;
    }

    std::vector<double> min_dists(num_items, 1e18);
    double max_d = 0.0;

    for (size_t i = 0; i < num_items; ++i) {
        double cx = centroids[i * 2];
        double cy = centroids[i * 2 + 1];

        for (size_t e = 0; e < num_epicenters; ++e) {
            double ex = epicenters[e * 2];
            double ey = epicenters[e * 2 + 1];
            double d = std::hypot(cx - ex, cy - ey);
            if (d < min_dists[i]) {
                min_dists[i] = d;
            }
        }
        if (min_dists[i] > max_d) max_d = min_dists[i];
    }

    if (max_d <= 0.0) max_d = 1.0;

    double travel_time = duration_frames * 0.75 * (speed > 0 ? speed : 1.0);
    double reveal_window = duration_frames * 0.25;

    for (size_t i = 0; i < num_items; ++i) {
        double norm_d = min_dists[i] / max_d;
        double start_f = norm_d * travel_time;
        out_start_frames[i] = start_f;
        out_end_frames[i] = start_f + reveal_window;
    }

    return 0;
}


// ============================================================================
// Sutherland-Hodgman Boolean Polygon Vector Clipping against BBox
// ============================================================================

struct Point2D { double x, y; };

static std::vector<Point2D> _clip_poly_single_edge(
    const std::vector<Point2D>& poly,
    int edge, double edge_val
) {
    std::vector<Point2D> out;
    if (poly.empty()) return out;

    auto is_inside = [edge, edge_val](const Point2D& p) -> bool {
        switch (edge) {
            case 0: return p.x >= edge_val; // Left: x >= minx
            case 1: return p.x <= edge_val; // Right: x <= maxx
            case 2: return p.y >= edge_val; // Bottom: y >= miny
            case 3: return p.y <= edge_val; // Top: y <= maxy
        }
        return true;
    };

    auto compute_intersect = [edge, edge_val](const Point2D& p1, const Point2D& p2) -> Point2D {
        Point2D res;
        double dx = p2.x - p1.x;
        double dy = p2.y - p1.y;
        switch (edge) {
            case 0: // x = minx
            case 1: // x = maxx
                res.x = edge_val;
                res.y = (std::abs(dx) > 1e-12) ? p1.y + dy * (edge_val - p1.x) / dx : p1.y;
                break;
            case 2: // y = miny
            case 3: // y = maxy
                res.y = edge_val;
                res.x = (std::abs(dy) > 1e-12) ? p1.x + dx * (edge_val - p1.y) / dy : p1.x;
                break;
        }
        return res;
    };

    Point2D prev = poly.back();
    bool prev_inside = is_inside(prev);

    for (const auto& curr : poly) {
        bool curr_inside = is_inside(curr);
        if (curr_inside) {
            if (!prev_inside) {
                out.push_back(compute_intersect(prev, curr));
            }
            out.push_back(curr);
        } else if (prev_inside) {
            out.push_back(compute_intersect(prev, curr));
        }
        prev = curr;
        prev_inside = curr_inside;
    }
    return out;
}

BL_GEO_EXPORT int nm_geo_clip_polygon_bbox(
    const double* in_coords, int64_t in_count,
    double minx, double miny, double maxx, double maxy,
    double* out_coords, int64_t* out_count
) {
    if (!in_coords || in_count < 3 || !out_coords || !out_count) return -1;

    std::vector<Point2D> poly(in_count);
    for (int64_t i = 0; i < in_count; ++i) {
        poly[i] = { in_coords[i * 2], in_coords[i * 2 + 1] };
    }

    std::vector<Point2D> clipped = _clip_poly_single_edge(poly, 0, minx);
    if (!clipped.empty()) clipped = _clip_poly_single_edge(clipped, 1, maxx);
    if (!clipped.empty()) clipped = _clip_poly_single_edge(clipped, 2, miny);
    if (!clipped.empty()) clipped = _clip_poly_single_edge(clipped, 3, maxy);

    *out_count = (int64_t)clipped.size();
    for (size_t i = 0; i < clipped.size(); ++i) {
        out_coords[i * 2] = clipped[i].x;
        out_coords[i * 2 + 1] = clipped[i].y;
    }
    return 0;
}

BL_GEO_EXPORT int bl_geo_build_polygon_rings(
    BLPathCore* path,
    const double* coords,
    const int64_t* ring_offsets,
    int64_t num_rings,
    const GeoTransform* gt
) {
    if (!path || !coords || !ring_offsets || num_rings <= 0) return -1;

    for (int64_t r = 0; r < num_rings; ++r) {
        int64_t start_idx = ring_offsets[r];
        int64_t end_idx = ring_offsets[r + 1];
        int64_t pt_count = end_idx - start_idx;
        if (pt_count < 3) continue;

        double sx, sy;
        transform_pt(gt, coords[start_idx * 2], coords[start_idx * 2 + 1], sx, sy);
        bl_path_move_to(path, sx, sy);

        for (int64_t i = start_idx + 1; i < end_idx; ++i) {
            transform_pt(gt, coords[i * 2], coords[i * 2 + 1], sx, sy);
            bl_path_line_to(path, sx, sy);
        }
        bl_path_close(path); // Essential: prevents spiderweb/fanning glitches!
    }
    return 0;
}

// ============================================================================
// Structure of Arrays (SoA) & Path Memory Arena
// ============================================================================

struct BLBox2D {
    double x0, y0, x1, y1;
};

struct NativeGeoLayerSoA {
    size_t count;
    size_t capacity;
    BLPathCore* paths;       // contiguous array of BLPathCore [capacity]
    BLBox2D* bboxes;         // contiguous array of BLBox2D [capacity]
    double* centroids;       // contiguous array of (cx, cy) [capacity * 2]
    float* distances;        // contiguous array of distances [capacity]
    double* start_frames;    // contiguous array of start_frames [capacity]
    double* end_frames;      // contiguous array of end_frames [capacity]
    uint8_t* is_active;      // contiguous array of active visibility flags [capacity]
    void* arena_memory;      // backing contiguous memory arena
    size_t arena_size;
};

BL_GEO_EXPORT NativeGeoLayerSoA* nm_geo_soa_create(size_t capacity) {
    if (capacity == 0) capacity = 1024;

    size_t sz_paths = capacity * sizeof(BLPathCore);
    size_t sz_bboxes = capacity * sizeof(BLBox2D);
    size_t sz_centroids = capacity * 2 * sizeof(double);
    size_t sz_distances = capacity * sizeof(float);
    size_t sz_start_frames = capacity * sizeof(double);
    size_t sz_end_frames = capacity * sizeof(double);
    size_t sz_active = capacity * sizeof(uint8_t);
    size_t total_alloc = sz_paths + sz_bboxes + sz_centroids + sz_distances + sz_start_frames + sz_end_frames + sz_active + 256;

    uint8_t* mem = (uint8_t*)malloc(total_alloc);
    if (!mem) return nullptr;
    memset(mem, 0, total_alloc);

    NativeGeoLayerSoA* layer = new NativeGeoLayerSoA();
    layer->count = 0;
    layer->capacity = capacity;
    layer->arena_memory = mem;
    layer->arena_size = total_alloc;

    uint8_t* ptr = mem;
    layer->paths = (BLPathCore*)ptr; ptr += sz_paths;
    layer->bboxes = (BLBox2D*)ptr; ptr += sz_bboxes;
    layer->centroids = (double*)ptr; ptr += sz_centroids;
    layer->distances = (float*)ptr; ptr += sz_distances;
    layer->start_frames = (double*)ptr; ptr += sz_start_frames;
    layer->end_frames = (double*)ptr; ptr += sz_end_frames;
    layer->is_active = (uint8_t*)ptr; ptr += sz_active;

    for (size_t i = 0; i < capacity; ++i) {
        bl_path_init(&layer->paths[i]);
        layer->is_active[i] = 1;
    }
    return layer;
}

BL_GEO_EXPORT void nm_geo_soa_destroy(NativeGeoLayerSoA* layer) {
    if (!layer) return;
    for (size_t i = 0; i < layer->capacity; ++i) {
        bl_path_destroy(&layer->paths[i]);
    }
    if (layer->arena_memory) {
        free(layer->arena_memory);
    }
    delete layer;
}

BL_GEO_EXPORT int nm_geo_soa_add_polygon(
    NativeGeoLayerSoA* layer,
    const double* coords,
    const int64_t* ring_offsets,
    int64_t num_rings,
    const GeoTransform* gt
) {
    if (!layer || layer->count >= layer->capacity || !coords || !ring_offsets || num_rings <= 0) {
        return -1;
    }

    size_t idx = layer->count;
    BLPathCore* path = &layer->paths[idx];
    bl_path_reset(path);

    double minx = 1e30, miny = 1e30, maxx = -1e30, maxy = -1e30;
    double sum_x = 0.0, sum_y = 0.0;
    int64_t total_pts = 0;

    for (int64_t r = 0; r < num_rings; ++r) {
        int64_t start_idx = ring_offsets[r];
        int64_t end_idx = ring_offsets[r + 1];
        int64_t pt_count = end_idx - start_idx;
        if (pt_count < 3) continue;

        double sx, sy;
        transform_pt(gt, coords[start_idx * 2], coords[start_idx * 2 + 1], sx, sy);
        bl_path_move_to(path, sx, sy);

        if (sx < minx) minx = sx;
        if (sx > maxx) maxx = sx;
        if (sy < miny) miny = sy;
        if (sy > maxy) maxy = sy;
        sum_x += sx;
        sum_y += sy;
        total_pts++;

        for (int64_t i = start_idx + 1; i < end_idx; ++i) {
            transform_pt(gt, coords[i * 2], coords[i * 2 + 1], sx, sy);
            bl_path_line_to(path, sx, sy);

            if (sx < minx) minx = sx;
            if (sx > maxx) maxx = sx;
            if (sy < miny) miny = sy;
            if (sy > maxy) maxy = sy;
            sum_x += sx;
            sum_y += sy;
            total_pts++;
        }
        bl_path_close(path); // Auto-close ring
    }

    if (total_pts == 0) {
        minx = miny = maxx = maxy = 0.0;
        sum_x = sum_y = 0.0;
        total_pts = 1;
    }

    layer->bboxes[idx].x0 = minx;
    layer->bboxes[idx].y0 = miny;
    layer->bboxes[idx].x1 = maxx;
    layer->bboxes[idx].y1 = maxy;

    layer->centroids[idx * 2] = sum_x / (double)total_pts;
    layer->centroids[idx * 2 + 1] = sum_y / (double)total_pts;

    layer->is_active[idx] = 1;
    layer->count++;
    return (int)idx;
}

BL_GEO_EXPORT int nm_geo_soa_from_wkb_batch(
    NativeGeoLayerSoA* layer,
    const uint8_t* const* wkb_ptrs,
    const size_t* wkb_sizes,
    size_t num_geoms,
    const GeoTransform* gt
) {
    if (!layer || !wkb_ptrs || !wkb_sizes || num_geoms == 0) return -1;

    for (size_t g = 0; g < num_geoms; ++g) {
        if (layer->count >= layer->capacity) break;
        if (!wkb_ptrs[g] || wkb_sizes[g] < 5) continue;

        WkbParsedGeometry geom;
        _parse_wkb_single(wkb_ptrs[g], wkb_sizes[g], geom);
        if (geom.num_points == 0) continue;

        size_t idx = layer->count;
        BLPathCore* path = &layer->paths[idx];
        bl_path_reset(path);

        double minx = 1e30, miny = 1e30, maxx = -1e30, maxy = -1e30;
        double sum_x = 0.0, sum_y = 0.0;
        int64_t total_pts = 0;

        if (geom.geom_type == 1) { // Point
            if (geom.coords.size() >= 2) {
                double sx, sy;
                transform_pt(gt, geom.coords[0], geom.coords[1], sx, sy);
                bl_path_move_to(path, sx - 2.0, sy);
                bl_path_line_to(path, sx + 2.0, sy);
                minx = maxx = sx; miny = maxy = sy;
                sum_x = sx; sum_y = sy; total_pts = 1;
            }
        } else if (geom.geom_type == 2) { // LineString
            int64_t n_pts = (int64_t)(geom.coords.size() / 2);
            if (n_pts >= 2) {
                double sx, sy;
                transform_pt(gt, geom.coords[0], geom.coords[1], sx, sy);
                bl_path_move_to(path, sx, sy);
                if (sx < minx) minx = sx; if (sx > maxx) maxx = sx;
                if (sy < miny) miny = sy; if (sy > maxy) maxy = sy;
                sum_x += sx; sum_y += sy; total_pts++;

                for (int64_t i = 1; i < n_pts; ++i) {
                    transform_pt(gt, geom.coords[i * 2], geom.coords[i * 2 + 1], sx, sy);
                    bl_path_line_to(path, sx, sy);
                    if (sx < minx) minx = sx; if (sx > maxx) maxx = sx;
                    if (sy < miny) miny = sy; if (sy > maxy) maxy = sy;
                    sum_x += sx; sum_y += sy; total_pts++;
                }
            }
        } else if (geom.geom_type == 3 || geom.geom_type == 6) { // Polygon or MultiPolygon
            size_t n_parts = geom.part_offsets.size();
            for (size_t p = 0; p < n_parts; ++p) {
                int64_t start_pt = geom.part_offsets[p];
                int64_t end_pt = (p + 1 < n_parts) ? geom.part_offsets[p + 1] : (int64_t)(geom.coords.size() / 2);
                int64_t pt_cnt = end_pt - start_pt;
                if (pt_cnt < 3) continue;

                double sx, sy;
                transform_pt(gt, geom.coords[start_pt * 2], geom.coords[start_pt * 2 + 1], sx, sy);
                bl_path_move_to(path, sx, sy);
                if (sx < minx) minx = sx; if (sx > maxx) maxx = sx;
                if (sy < miny) miny = sy; if (sy > maxy) maxy = sy;
                sum_x += sx; sum_y += sy; total_pts++;

                for (int64_t i = start_pt + 1; i < end_pt; ++i) {
                    transform_pt(gt, geom.coords[i * 2], geom.coords[i * 2 + 1], sx, sy);
                    bl_path_line_to(path, sx, sy);
                    if (sx < minx) minx = sx; if (sx > maxx) maxx = sx;
                    if (sy < miny) miny = sy; if (sy > maxy) maxy = sy;
                    sum_x += sx; sum_y += sy; total_pts++;
                }
                bl_path_close(path); // Auto-close ring: cuts interior holes with EVEN_ODD!
            }
        } else if (geom.geom_type == 5) { // MultiLineString
            size_t n_parts = geom.part_offsets.size();
            for (size_t p = 0; p < n_parts; ++p) {
                int64_t start_pt = geom.part_offsets[p];
                int64_t end_pt = (p + 1 < n_parts) ? geom.part_offsets[p + 1] : (int64_t)(geom.coords.size() / 2);
                int64_t pt_cnt = end_pt - start_pt;
                if (pt_cnt < 2) continue;

                double sx, sy;
                transform_pt(gt, geom.coords[start_pt * 2], geom.coords[start_pt * 2 + 1], sx, sy);
                bl_path_move_to(path, sx, sy);
                if (sx < minx) minx = sx; if (sx > maxx) maxx = sx;
                if (sy < miny) miny = sy; if (sy > maxy) maxy = sy;
                sum_x += sx; sum_y += sy; total_pts++;

                for (int64_t i = start_pt + 1; i < end_pt; ++i) {
                    transform_pt(gt, geom.coords[i * 2], geom.coords[i * 2 + 1], sx, sy);
                    bl_path_line_to(path, sx, sy);
                    if (sx < minx) minx = sx; if (sx > maxx) maxx = sx;
                    if (sy < miny) miny = sy; if (sy > maxy) maxy = sy;
                    sum_x += sx; sum_y += sy; total_pts++;
                }
            }
        }

        if (total_pts == 0) {
            minx = miny = maxx = maxy = 0.0;
            sum_x = sum_y = 0.0;
            total_pts = 1;
        }

        layer->bboxes[idx].x0 = minx;
        layer->bboxes[idx].y0 = miny;
        layer->bboxes[idx].x1 = maxx;
        layer->bboxes[idx].y1 = maxy;

        layer->centroids[idx * 2] = sum_x / (double)total_pts;
        layer->centroids[idx * 2 + 1] = sum_y / (double)total_pts;

        layer->is_active[idx] = 1;
        layer->count++;
    }

    return (int)layer->count;
}

BL_GEO_EXPORT int nm_geo_soa_compute_propagation(
    NativeGeoLayerSoA* layer,
    const double* epicenters,
    size_t num_epicenters,
    double duration_frames,
    double speed
) {
    if (!layer || layer->count == 0 || !epicenters || num_epicenters == 0) return -1;

    size_t n = layer->count;
    double max_d = 0.0;

    for (size_t i = 0; i < n; ++i) {
        double cx = layer->centroids[i * 2];
        double cy = layer->centroids[i * 2 + 1];
        double min_d = 1e18;

        for (size_t e = 0; e < num_epicenters; ++e) {
            double ex = epicenters[e * 2];
            double ey = epicenters[e * 2 + 1];
            double d = std::hypot(cx - ex, cy - ey);
            if (d < min_d) min_d = d;
        }

        layer->distances[i] = (float)min_d;
        if (min_d > max_d) max_d = min_d;
    }

    if (max_d <= 0.0) max_d = 1.0;

    double travel_time = duration_frames * 0.75 * (speed > 0.0 ? speed : 1.0);
    double reveal_window = duration_frames * 0.25;

    for (size_t i = 0; i < n; ++i) {
        double norm_d = (double)layer->distances[i] / max_d;
        double sf = norm_d * travel_time;
        layer->start_frames[i] = sf;
        layer->end_frames[i] = sf + reveal_window;
    }

    return 0;
}

BL_GEO_EXPORT int nm_geo_soa_cull_and_update(
    NativeGeoLayerSoA* layer,
    double clip_x0, double clip_y0, double clip_x1, double clip_y1,
    double current_frame,
    int64_t* out_visible_count
) {
    if (!layer) return -1;
    size_t n = layer->count;
    int64_t vis_count = 0;

    for (size_t i = 0; i < n; ++i) {
        if (current_frame >= 0.0 && current_frame < layer->start_frames[i]) {
            layer->is_active[i] = 0;
            continue;
        }

        const BLBox2D& b = layer->bboxes[i];
        if (b.x1 < clip_x0 || b.x0 > clip_x1 || b.y1 < clip_y0 || b.y0 > clip_y1) {
            layer->is_active[i] = 0;
            continue;
        }

        layer->is_active[i] = 1;
        vis_count++;
    }

    if (out_visible_count) *out_visible_count = vis_count;
    return 0;
}

BL_GEO_EXPORT int nm_geo_soa_render_dual_buffer(
    void* ctx_ptr,
    const NativeGeoLayerSoA* layer,
    uint32_t fill_rgba,
    uint32_t stroke_rgba,
    double stroke_width,
    int fill_enabled,
    int stroke_enabled
) {
    if (!ctx_ptr || !layer || layer->count == 0) return -1;
    BLContextCore* ctx = (BLContextCore*)ctx_ptr;
    size_t n = layer->count;

    bl_context_set_fill_rule(ctx, 1 /* EVEN_ODD */);

    // Pass 1: Batch Fill Pass (all fills in a tight unrolled loop)
    if (fill_enabled) {
        for (size_t i = 0; i < n; ++i) {
            if (layer->is_active[i]) {
                bl_context_fill_geometry_rgba32(ctx, 21 /* BL_GEOMETRY_TYPE_PATH */, &layer->paths[i], fill_rgba);
            }
        }
    }

    // Pass 2: Batch Stroke Pass (all strokes in a tight loop)
    if (stroke_enabled && stroke_width > 0.0) {
        bl_context_set_stroke_width(ctx, stroke_width);
        for (size_t i = 0; i < n; ++i) {
            if (layer->is_active[i]) {
                bl_context_stroke_geometry_rgba32(ctx, 21 /* BL_GEOMETRY_TYPE_PATH */, &layer->paths[i], stroke_rgba);
            }
        }
    }

    return 0;
}

static inline double _geo_easing_back_out(double t) {
    if (t <= 0.0) return 0.0;
    if (t >= 1.0) return 1.0;
    double c1 = 1.70158;
    double c3 = c1 + 1.0;
    double t_m_1 = t - 1.0;
    return 1.0 + c3 * (t_m_1 * t_m_1 * t_m_1) + c1 * (t_m_1 * t_m_1);
}

BL_GEO_EXPORT int nm_geo_soa_render_popin(
    void* ctx_ptr,
    const NativeGeoLayerSoA* layer,
    double current_frame,
    uint32_t fill_rgba,
    uint32_t stroke_rgba,
    double stroke_width,
    int fill_enabled,
    int stroke_enabled
) {
    if (!ctx_ptr || !layer || layer->count == 0) return -1;
    BLContextCore* ctx = (BLContextCore*)ctx_ptr;
    size_t n = layer->count;

    bl_context_set_fill_rule(ctx, 1 /* EVEN_ODD */);

    // Pass 1: Batch Fill Pass with Centroid Spring Pop-In
    if (fill_enabled) {
        for (size_t i = 0; i < n; ++i) {
            if (!layer->is_active[i]) continue;
            double sf = layer->start_frames[i];
            double ef = layer->end_frames[i];

            if (current_frame < sf) continue;

            if (current_frame >= ef) {
                bl_context_fill_geometry_rgba32(ctx, 21, &layer->paths[i], fill_rgba);
            } else {
                double t = (current_frame - sf) / (ef - sf);
                double scale = _geo_easing_back_out(t);
                if (scale > 0.001) {
                    double cx = layer->centroids[i * 2];
                    double cy = layer->centroids[i * 2 + 1];
                    BLMatrix2D m;
                    m.m00 = scale;
                    m.m01 = 0.0;
                    m.m10 = 0.0;
                    m.m11 = scale;
                    m.m20 = cx * (1.0 - scale);
                    m.m21 = cy * (1.0 - scale);

                    bl_context_apply_transform_op(ctx, 1 /* ASSIGN */, &m);
                    bl_context_fill_geometry_rgba32(ctx, 21, &layer->paths[i], fill_rgba);
                    bl_context_apply_transform_op(ctx, 0 /* RESET */, NULL);
                }
            }
        }
    }

    // Pass 2: Batch Stroke Pass with Centroid Spring Pop-In
    if (stroke_enabled && stroke_width > 0.0) {
        bl_context_set_stroke_width(ctx, stroke_width);
        for (size_t i = 0; i < n; ++i) {
            if (!layer->is_active[i]) continue;
            double sf = layer->start_frames[i];
            double ef = layer->end_frames[i];

            if (current_frame < sf) continue;

            if (current_frame >= ef) {
                bl_context_stroke_geometry_rgba32(ctx, 21, &layer->paths[i], stroke_rgba);
            } else {
                double t = (current_frame - sf) / (ef - sf);
                double scale = _geo_easing_back_out(t);
                if (scale > 0.001) {
                    double cx = layer->centroids[i * 2];
                    double cy = layer->centroids[i * 2 + 1];
                    BLMatrix2D m;
                    m.m00 = scale;
                    m.m01 = 0.0;
                    m.m10 = 0.0;
                    m.m11 = scale;
                    m.m20 = cx * (1.0 - scale);
                    m.m21 = cy * (1.0 - scale);

                    bl_context_apply_transform_op(ctx, 1 /* ASSIGN */, &m);
                    bl_context_stroke_geometry_rgba32(ctx, 21, &layer->paths[i], stroke_rgba);
                    bl_context_apply_transform_op(ctx, 0 /* RESET */, NULL);
                }
            }
        }
    }

    return 0;
}

// ============================================================================
// Flat Uniform Grid Spatial Indexing
// ============================================================================

struct UniformGrid2D {
    double minx, miny, maxx, maxy;
    double cell_w, cell_h;
    int32_t grid_w, grid_h;
    int32_t* cell_counts;     // [grid_w * grid_h]
    int32_t* cell_offsets;    // [grid_w * grid_h + 1]
    int32_t* item_indices;    // [total_references]
    int32_t total_references;
};

BL_GEO_EXPORT UniformGrid2D* nm_geo_grid_create(
    double minx, double miny, double maxx, double maxy,
    int32_t grid_w, int32_t grid_h
) {
    if (grid_w <= 0) grid_w = 32;
    if (grid_h <= 0) grid_h = 32;
    if (maxx <= minx) maxx = minx + 1.0;
    if (maxy <= miny) maxy = miny + 1.0;

    UniformGrid2D* grid = new UniformGrid2D();
    grid->minx = minx;
    grid->miny = miny;
    grid->maxx = maxx;
    grid->maxy = maxy;
    grid->grid_w = grid_w;
    grid->grid_h = grid_h;
    grid->cell_w = (maxx - minx) / (double)grid_w;
    grid->cell_h = (maxy - miny) / (double)grid_h;

    size_t num_cells = (size_t)(grid_w * grid_h);
    grid->cell_counts = (int32_t*)calloc(num_cells, sizeof(int32_t));
    grid->cell_offsets = (int32_t*)calloc(num_cells + 1, sizeof(int32_t));
    grid->item_indices = nullptr;
    grid->total_references = 0;

    return grid;
}

BL_GEO_EXPORT int nm_geo_grid_build(
    UniformGrid2D* grid,
    const BLBox2D* bboxes,
    size_t count
) {
    if (!grid || !bboxes || count == 0) return -1;

    int32_t gw = grid->grid_w;
    int32_t gh = grid->grid_h;
    int32_t num_cells = gw * gh;

    memset(grid->cell_counts, 0, num_cells * sizeof(int32_t));

    auto to_cx = [grid, gw](double x) -> int32_t {
        int32_t c = (int32_t)((x - grid->minx) / grid->cell_w);
        if (c < 0) c = 0;
        if (c >= gw) c = gw - 1;
        return c;
    };
    auto to_cy = [grid, gh](double y) -> int32_t {
        int32_t c = (int32_t)((y - grid->miny) / grid->cell_h);
        if (c < 0) c = 0;
        if (c >= gh) c = gh - 1;
        return c;
    };

    int32_t total_refs = 0;
    for (size_t i = 0; i < count; ++i) {
        int32_t cx0 = to_cx(bboxes[i].x0);
        int32_t cx1 = to_cx(bboxes[i].x1);
        int32_t cy0 = to_cy(bboxes[i].y0);
        int32_t cy1 = to_cy(bboxes[i].y1);

        for (int32_t y = cy0; y <= cy1; ++y) {
            for (int32_t x = cx0; x <= cx1; ++x) {
                int32_t cell = y * gw + x;
                grid->cell_counts[cell]++;
                total_refs++;
            }
        }
    }

    grid->total_references = total_refs;
    if (grid->item_indices) free(grid->item_indices);
    grid->item_indices = (int32_t*)malloc(total_refs * sizeof(int32_t));

    int32_t cur_offset = 0;
    for (int32_t c = 0; c < num_cells; ++c) {
        grid->cell_offsets[c] = cur_offset;
        cur_offset += grid->cell_counts[c];
    }
    grid->cell_offsets[num_cells] = cur_offset;

    memset(grid->cell_counts, 0, num_cells * sizeof(int32_t));

    for (size_t i = 0; i < count; ++i) {
        int32_t cx0 = to_cx(bboxes[i].x0);
        int32_t cx1 = to_cx(bboxes[i].x1);
        int32_t cy0 = to_cy(bboxes[i].y0);
        int32_t cy1 = to_cy(bboxes[i].y1);

        for (int32_t y = cy0; y <= cy1; ++y) {
            for (int32_t x = cx0; x <= cx1; ++x) {
                int32_t cell = y * gw + x;
                int32_t write_pos = grid->cell_offsets[cell] + grid->cell_counts[cell];
                grid->item_indices[write_pos] = (int32_t)i;
                grid->cell_counts[cell]++;
            }
        }
    }

    return 0;
}

BL_GEO_EXPORT int nm_geo_grid_query_bbox(
    const UniformGrid2D* grid,
    double qx0, double qy0, double qx1, double qy1,
    int32_t* out_indices,
    int32_t max_indices,
    int32_t* out_count
) {
    if (!grid || !out_indices || max_indices <= 0 || !out_count) return -1;

    int32_t gw = grid->grid_w;
    int32_t gh = grid->grid_h;

    auto to_cx = [grid, gw](double x) -> int32_t {
        int32_t c = (int32_t)((x - grid->minx) / grid->cell_w);
        if (c < 0) c = 0;
        if (c >= gw) c = gw - 1;
        return c;
    };
    auto to_cy = [grid, gh](double y) -> int32_t {
        int32_t c = (int32_t)((y - grid->miny) / grid->cell_h);
        if (c < 0) c = 0;
        if (c >= gh) c = gh - 1;
        return c;
    };

    int32_t cx0 = to_cx(qx0);
    int32_t cx1 = to_cx(qx1);
    int32_t cy0 = to_cy(qy0);
    int32_t cy1 = to_cy(qy1);

    int32_t found = 0;
    int32_t max_item_id = -1;

    for (int32_t y = cy0; y <= cy1; ++y) {
        for (int32_t x = cx0; x <= cx1; ++x) {
            int32_t cell = y * gw + x;
            int32_t cnt = grid->cell_counts[cell];
            int32_t off = grid->cell_offsets[cell];
            for (int32_t k = 0; k < cnt; ++k) {
                int32_t idx = grid->item_indices[off + k];
                if (idx > max_item_id) max_item_id = idx;
            }
        }
    }

    if (max_item_id < 0) {
        *out_count = 0;
        return 0;
    }

    std::vector<bool> visited(max_item_id + 1, false);

    for (int32_t y = cy0; y <= cy1 && found < max_indices; ++y) {
        for (int32_t x = cx0; x <= cx1 && found < max_indices; ++x) {
            int32_t cell = y * gw + x;
            int32_t cnt = grid->cell_counts[cell];
            int32_t off = grid->cell_offsets[cell];
            for (int32_t k = 0; k < cnt && found < max_indices; ++k) {
                int32_t idx = grid->item_indices[off + k];
                if (!visited[idx]) {
                    visited[idx] = true;
                    out_indices[found++] = idx;
                }
            }
        }
    }

    *out_count = found;
    return 0;
}

BL_GEO_EXPORT void nm_geo_grid_destroy(UniformGrid2D* grid) {
    if (!grid) return;
    if (grid->cell_counts) free(grid->cell_counts);
    if (grid->cell_offsets) free(grid->cell_offsets);
    if (grid->item_indices) free(grid->item_indices);
    delete grid;
}

} // extern "C"