// gpr_poc —— POC-B headless 度量 CLI。
//
// 串起整条地基：发现 14 通道 .iprb + .ord → assembleGprSurvey → buildGprVolume
// → ChunkedVolumeStore::write → buildPyramid → WholeVolumeSource(load)，
// 在真实/合成数据上输出可测的真实指标（耗时/维度/体积/压缩比/加载/峰值内存）。
//
// 子命令：
//   gpr_poc build <dir> [--line 001] [--cellXY 0.2] [--cellZ 0.05] [--out <storeDir>] [--levels 2]
//   gpr_poc load  <storeDir>
//   gpr_poc selftest
//   gpr_poc offscreen-smoke               —— 离屏 GL 闸门冒烟
//   gpr_poc renderB <storeDir> [--frames 120]  —— 离屏体绘制/切片 fps 基准

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

#include "Probe.hpp"

#include "core/algo/GprVolumeBuilder.hpp"
#include "core/algo/IInterpolator.hpp"
#include "core/model/GprSurvey.hpp"
#include "core/model/ColorScale.hpp"
#include "core/model/ScalarVolumeI16.hpp"
#include "data/store/ChunkedVolumeStore.hpp"
#include "io/gpr/GprSurveyAssembler.hpp"
#include "render/actors/VoxelActor.hpp"
#include "render/source/OutOfCoreSource.hpp"
#include "render/source/WholeVolumeSource.hpp"

// ---- VTK 离屏渲染 ----
#include <vtkActor.h>
#include <vtkCamera.h>
#include <vtkCubeSource.h>
#include <vtkGPUVolumeRayCastMapper.h>
#include <vtkOpenGLGPUVolumeRayCastMapper.h>
#include <vtkImageActor.h>
#include <vtkImageData.h>
#include <vtkImageMapToColors.h>
#include <vtkImageMapper3D.h>
#include <vtkImageReslice.h>
#include <vtkLookupTable.h>
#include <vtkColorTransferFunction.h>
#include <vtkMultiBlockDataSet.h>
#include <vtkMultiBlockVolumeMapper.h>
#include <vtkNew.h>
#include <vtkPiecewiseFunction.h>
#include <vtkVolumeProperty.h>
#include <vtkObjectFactory.h>
#include <vtkOutputWindow.h>
#include <vtkPNGWriter.h>
#include <vtkPointData.h>
#include <vtkPolyDataMapper.h>
#include <vtkProperty.h>
#include <vtkRenderWindow.h>
#include <vtkRenderer.h>
#include <vtkShortArray.h>
#include <vtkSmartPointer.h>
#include <vtkSmartVolumeMapper.h>
#include <vtkUnsignedCharArray.h>
#include <vtkVolume.h>
#include <vtkWindowToImageFilter.h>

namespace fs = std::filesystem;
using geopro::tools::Probe;
using geopro::tools::Stopwatch;

namespace {

constexpr int kChannels = 14;

// ---- 命令行参数解析（极简 --key value）----
struct Args {
  std::map<std::string, std::string> kv;
  std::vector<std::string> positional;

  std::string get(const std::string& key, const std::string& def) const {
    auto it = kv.find(key);
    return it != kv.end() ? it->second : def;
  }
};

Args parseArgs(int argc, char** argv, int start) {
  Args a;
  for (int i = start; i < argc; ++i) {
    std::string tok = argv[i];
    if (tok.rfind("--", 0) == 0 && i + 1 < argc) {
      a.kv[tok.substr(2)] = argv[i + 1];
      ++i;
    } else {
      a.positional.push_back(tok);
    }
  }
  return a;
}

// 把一行指标追加写入 last-metrics.txt（与可执行同目录的工具源目录无关，
// 写到当前工作目录便于汇总；CSV 风格一行）。
void writeMetricLine(const std::string& line) {
  std::ofstream f("last-metrics.txt", std::ios::app);
  if (f) f << line << "\n";
}

// 发现某线 14 通道 .iprb（按通道号 A01..A14 排序）+ 该线 .ord。
struct LineFiles {
  std::vector<std::string> iprb;  // 已按通道号排序
  std::string ord;
};

LineFiles discoverLine(const std::string& dir, const std::string& line) {
  LineFiles lf;
  std::map<int, std::string> byChannel;  // 通道号 → 路径（自动按号排序）
  std::string ordPath;

  for (const auto& e : fs::directory_iterator(dir)) {
    if (!e.is_regular_file()) continue;
    const std::string name = e.path().filename().string();
    const std::string ext = e.path().extension().string();

    // .ord：优先匹配本线（含 "_<line>" 且以 .ord 结尾），否则记下工区任一 .ord 作兜底。
    if (ext == ".ord") {
      if (name.find("_" + line + ".") != std::string::npos) {
        ordPath = e.path().string();
      } else if (ordPath.empty()) {
        ordPath = e.path().string();
      }
      continue;
    }

    // .iprb：匹配 "*_<line>_A<NN>.iprb"。
    if (ext != ".iprb") continue;
    const std::string tag = "_" + line + "_A";
    const std::size_t pos = name.find(tag);
    if (pos == std::string::npos) continue;
    const std::size_t numStart = pos + tag.size();
    std::size_t numEnd = numStart;
    while (numEnd < name.size() &&
           std::isdigit(static_cast<unsigned char>(name[numEnd]))) {
      ++numEnd;
    }
    if (numEnd == numStart) continue;
    const int ch = std::stoi(name.substr(numStart, numEnd - numStart));
    byChannel[ch] = e.path().string();
  }

  for (const auto& [ch, path] : byChannel) lf.iprb.push_back(path);
  lf.ord = ordPath;
  return lf;
}

// 由 survey 推 GridSpec：X 沿测线，Y 跨通道，Z 深度。
geopro::core::GridSpec specFromSurvey(const geopro::core::GprSurvey& s,
                                      double cellXY, double cellZ) {
  geopro::core::GridSpec spec{};

  const double rangeX =
      (s.ntraces > 1) ? (s.ntraces - 1) * s.dx : 0.0;
  const double y0 = s.channelY.empty() ? 0.0 : s.channelY.front();
  const double y1 = s.channelY.empty() ? 0.0 : s.channelY.back();
  const double rangeY = y1 - y0;
  const double rangeZ =
      (s.samples > 1) ? (s.samples - 1) * s.dz : 0.0;

  auto cells = [](double range, double cell) {
    if (cell <= 0.0) return 1;
    return static_cast<int>(std::ceil(range / cell)) + 1;
  };

  spec.ox = s.x0;
  spec.oy = y0;
  spec.oz = s.z0;
  spec.dx = cellXY;
  spec.dy = cellXY;
  spec.dz = cellZ;
  spec.nx = cells(rangeX, cellXY);
  spec.ny = cells(rangeY, cellXY);
  spec.nz = cells(rangeZ, cellZ);
  spec.power = 2.0;
  spec.maxDist = cellXY * 2.0;
  return spec;
}

// 落盘 data.bin 体积（所有 data*.bin 之和，含金字塔各级）。
std::int64_t storeDataBytes(const std::string& dir) {
  std::int64_t total = 0;
  for (const auto& e : fs::directory_iterator(dir)) {
    if (!e.is_regular_file()) continue;
    const std::string name = e.path().filename().string();
    if (name.rfind("data", 0) == 0 &&
        e.path().extension().string() == ".bin") {
      total += static_cast<std::int64_t>(e.file_size());
    }
  }
  return total;
}

int cmdBuild(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr << "用法: gpr_poc build <dir> [--line 001] [--cellXY 0.2] "
                 "[--cellZ 0.05] [--out <storeDir>] [--levels 2]\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  const std::string line = a.get("line", "001");
  const double cellXY = std::stod(a.get("cellXY", "0.2"));
  const double cellZ = std::stod(a.get("cellZ", "0.05"));
  const int levels = std::stoi(a.get("levels", "2"));
  const std::string out =
      a.get("out", (fs::temp_directory_path() / ("gpr_store_" + line)).string());

  std::cout << "[build] dir=" << dir << " line=" << line
            << " cellXY=" << cellXY << " cellZ=" << cellZ
            << " levels=" << levels << " out=" << out << "\n";

  const LineFiles lf = discoverLine(dir, line);
  std::cout << "[build] 发现通道数=" << lf.iprb.size()
            << " ord=" << (lf.ord.empty() ? "(无)" : lf.ord) << "\n";
  if (lf.iprb.size() != static_cast<std::size_t>(kChannels)) {
    std::cerr << "[build] 警告: 通道数 != " << kChannels
              << "（仍按发现数继续）\n";
  }
  if (lf.iprb.empty() || lf.ord.empty()) {
    std::cerr << "[build] 错误: 未发现 .iprb 或 .ord\n";
    return 1;
  }

  // 1) 装配
  Stopwatch swAsm;
  geopro::core::GprSurvey survey =
      geopro::io::gpr::assembleGprSurvey(lf.iprb, lf.ord);
  const double asmMs = swAsm.elapsedMs();
  std::cout << "[build] 装配完成 ntraces=" << survey.ntraces
            << " samples=" << survey.samples
            << " channels=" << survey.channelY.size()
            << " dx=" << survey.dx << " dz=" << survey.dz << "\n";

  // 2) 建体
  const geopro::core::GridSpec spec = specFromSurvey(survey, cellXY, cellZ);
  std::cout << "[build] GridSpec nx=" << spec.nx << " ny=" << spec.ny
            << " nz=" << spec.nz << " dx=" << spec.dx << " dy=" << spec.dy
            << " dz=" << spec.dz << " maxDist=" << spec.maxDist << "\n";

  Stopwatch swBuild;
  geopro::core::BuiltI16 built = geopro::core::buildGprVolume(survey, spec);
  const double buildMs = swBuild.elapsedMs();

  const std::int64_t nx = built.vol.nx(), ny = built.vol.ny(), nz = built.vol.nz();
  const std::int64_t voxels = nx * ny * nz;
  const std::int64_t rawBytes = voxels * 2;  // int16

  // 3) 落盘 + 金字塔
  fs::create_directories(out);
  Stopwatch swWrite;
  geopro::data::ChunkedVolumeStore::write(out, built);
  const double writeMs = swWrite.elapsedMs();

  Stopwatch swPyr;
  {
    geopro::data::ChunkedVolumeStore store(out);
    store.buildPyramid(levels);
  }
  const double pyrMs = swPyr.elapsedMs();

  const std::int64_t dataBytes = storeDataBytes(out);
  const double ratio =
      dataBytes > 0 ? static_cast<double>(rawBytes) / dataBytes : 0.0;
  const double peak = Probe::peakMemMB();

  std::cout << "\n=== build 指标 ===\n";
  std::cout << "装配耗时(ms)   : " << asmMs << "\n";
  std::cout << "建体耗时(ms)   : " << buildMs << "\n";
  std::cout << "落盘耗时(ms)   : " << writeMs << "\n";
  std::cout << "金字塔耗时(ms) : " << pyrMs << "\n";
  std::cout << "体维度         : " << nx << " x " << ny << " x " << nz << "\n";
  std::cout << "体素数         : " << voxels << "\n";
  std::cout << "原始体积(B)    : " << rawBytes << "  ("
            << rawBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "data.bin(B)    : " << dataBytes << "  ("
            << dataBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "压缩比         : " << ratio << " x\n";
  std::cout << "峰值内存(MB)   : " << peak << "\n";

  writeMetricLine(
      "build,line=" + line + ",cellXY=" + std::to_string(cellXY) +
      ",cellZ=" + std::to_string(cellZ) + ",nx=" + std::to_string(nx) +
      ",ny=" + std::to_string(ny) + ",nz=" + std::to_string(nz) +
      ",voxels=" + std::to_string(voxels) +
      ",rawB=" + std::to_string(rawBytes) +
      ",dataB=" + std::to_string(dataBytes) +
      ",ratio=" + std::to_string(ratio) + ",asmMs=" + std::to_string(asmMs) +
      ",buildMs=" + std::to_string(buildMs) +
      ",writeMs=" + std::to_string(writeMs) +
      ",pyrMs=" + std::to_string(pyrMs) + ",peakMB=" + std::to_string(peak));
  return 0;
}

int cmdLoad(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr << "用法: gpr_poc load <storeDir>\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  std::cout << "[load] storeDir=" << dir << "\n";

  Stopwatch sw;
  geopro::render::WholeVolumeSource src(dir);
  const double loadMs = sw.elapsedMs();

  const auto& m = src.meta();
  const std::int64_t voxels =
      static_cast<std::int64_t>(m.nx) * m.ny * m.nz;
  const std::int64_t wholeBytes = voxels * 2;  // VTK_SHORT
  const double peak = Probe::peakMemMB();

  std::cout << "\n=== load 指标 ===\n";
  std::cout << "加载耗时(ms) : " << loadMs << "\n";
  std::cout << "整卷维度     : " << m.nx << " x " << m.ny << " x " << m.nz
            << "\n";
  std::cout << "整卷字节(B)  : " << wholeBytes << "  ("
            << wholeBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "峰值内存(MB) : " << peak << "\n";

  writeMetricLine("load,dir=" + dir + ",nx=" + std::to_string(m.nx) +
                  ",ny=" + std::to_string(m.ny) +
                  ",nz=" + std::to_string(m.nz) +
                  ",wholeB=" + std::to_string(wholeBytes) +
                  ",loadMs=" + std::to_string(loadMs) +
                  ",peakMB=" + std::to_string(peak));
  return 0;
}

// ---- selftest：合成极小数据走完整 build→load 管线 ----

// 写一个极小通道的 .iprb + .iprh（samples 采样、traces 道，值 = base + t + s）。
void writeSyntheticChannel(const fs::path& iprbPath, int samples, int traces,
                           std::int16_t base) {
  const fs::path iprhPath =
      fs::path(iprbPath).replace_extension(".iprh");
  std::ofstream h(iprhPath);
  h << "SAMPLES: " << samples << "\n";
  h << "LAST TRACE: " << (traces - 1) << "\n";
  h << "CHANNELS: 2\n";
  h << "TIMEWINDOW: 100.0\n";
  h << "SOIL VELOCITY: 100.0\n";       // m/µs → ×1e6 → 1e8 m/s
  h << "DISTANCE INTERVAL: 0.05\n";
  h.close();

  std::ofstream b(iprbPath, std::ios::binary);
  // 布局 [trace*samples + s]，s 最快。
  for (int t = 0; t < traces; ++t) {
    for (int s = 0; s < samples; ++s) {
      const std::int16_t v =
          static_cast<std::int16_t>(base + t + s);
      b.write(reinterpret_cast<const char*>(&v), sizeof(v));
    }
  }
}

int cmdSelftest() {
  std::cout << "[selftest] 构造极小合成 survey（2 通道）...\n";
  const fs::path tmp =
      fs::temp_directory_path() / "gpr_poc_selftest";
  std::error_code ec;
  fs::remove_all(tmp, ec);
  fs::create_directories(tmp);

  const int samples = 8;
  const int traces = 12;

  // 2 通道 .iprb/.iprh + .ord（末列==1 标记有效通道，第 2 列为横偏 Y）。
  writeSyntheticChannel(tmp / "syn_001_A01.iprb", samples, traces,
                        /*base=*/100);
  writeSyntheticChannel(tmp / "syn_001_A02.iprb", samples, traces,
                        /*base=*/200);
  {
    std::ofstream ord(tmp / "syn_001.ord");
    ord << "0 0.000000 -1.5 1\n";
    ord << "1 1.000000 -1.5 1\n";
  }

  const std::vector<std::string> iprb = {
      (tmp / "syn_001_A01.iprb").string(),
      (tmp / "syn_001_A02.iprb").string()};
  const std::string ord = (tmp / "syn_001.ord").string();

  bool ok = true;
  auto check = [&](bool cond, const std::string& msg) {
    if (!cond) {
      std::cerr << "[selftest] FAIL: " << msg << "\n";
      ok = false;
    }
  };

  try {
    // 装配
    geopro::core::GprSurvey survey =
        geopro::io::gpr::assembleGprSurvey(iprb, ord);
    check(survey.ntraces == traces, "ntraces");
    check(survey.samples == samples, "samples");
    check(survey.channelY.size() == 2, "channels");
    // channelY 升序：A01 偏移 0.0 在前，A02 偏移 1.0 在后。
    check(survey.channelY.front() < survey.channelY.back(), "channelY 升序");

    // 建体：cellXY 取通道间距 1.0 → ny=2；cellZ 较细确保 nz>1。
    const double cellXY = 1.0;
    const double cellZ = std::max(survey.dz, 1e-12);
    const geopro::core::GridSpec spec =
        specFromSurvey(survey, cellXY, cellZ);
    std::cout << "[selftest] GridSpec " << spec.nx << "x" << spec.ny << "x"
              << spec.nz << " dz=" << spec.dz << "\n";
    check(spec.ny == 2, "ny==2");

    geopro::core::BuiltI16 built =
        geopro::core::buildGprVolume(survey, spec);
    check(built.vol.nx() == spec.nx, "built nx");
    check(built.vol.ny() == spec.ny, "built ny");
    check(built.vol.nz() == spec.nz, "built nz");

    // 落盘 + 金字塔
    const std::string store = (tmp / "store").string();
    fs::create_directories(store);
    geopro::data::ChunkedVolumeStore::write(store, built, /*brick=*/4);
    {
      geopro::data::ChunkedVolumeStore s(store);
      s.buildPyramid(1);
      check(s.levels() == 2, "金字塔层数==2");
    }

    // 加载整卷，校验维度一致
    geopro::render::WholeVolumeSource src(store);
    check(src.meta().nx == spec.nx, "load nx");
    check(src.meta().ny == spec.ny, "load ny");
    check(src.meta().nz == spec.nz, "load nz");

    // 某体素值合理性：x0/y0 角点应有非 blank 量化值（落格命中首道首通道）。
    const std::int16_t q = built.vol.at(0, 0, 0);
    check(q != geopro::core::ScalarVolumeI16::kBlank, "(0,0,0) 非 blank");
  } catch (const std::exception& e) {
    std::cerr << "[selftest] 异常: " << e.what() << "\n";
    ok = false;
  }

  fs::remove_all(tmp, ec);
  std::cout << "[selftest] " << (ok ? "PASS" : "FAIL") << "\n";
  return ok ? 0 : 1;
}

// ============================================================================
// 离屏 GPU 渲染基准（POC-B）
// ============================================================================

// 捕获 VTK 错误输出的 OutputWindow：用于侦测体绘制时 vtkVolumeTexture 报的
// "Invalid texture dimensions" / "MAX_3D_TEXTURE_SIZE" —— 一旦出现，说明整卷
// 单张 3D 纹理上传失败，体绘制 fps 无意义，必须如实标 INVALID（绝不当真上报）。
class CapturingOutputWindow : public vtkOutputWindow {
 public:
  static CapturingOutputWindow* New();
  vtkTypeMacro(CapturingOutputWindow, vtkOutputWindow);

  void DisplayText(const char* txt) override {
    if (txt) {
      const std::string s(txt);
      captured_ += s;
      if (s.find("texture dimensions") != std::string::npos ||
          s.find("MAX_3D_TEXTURE_SIZE") != std::string::npos) {
        textureError_ = true;
      }
    }
    // 仍透传到 stderr，便于人工查看。
    if (txt) std::cerr << txt;
  }

  bool textureError() const { return textureError_; }
  const std::string& captured() const { return captured_; }

 private:
  std::string captured_;
  bool textureError_ = false;
};
vtkStandardNewMacro(CapturingOutputWindow);

// 创建一个离屏 vtkRenderWindow（VTK9.6：SetShowWindow(false)+OffScreenRenderingOn）。
vtkSmartPointer<vtkRenderWindow> makeOffscreenWindow(int w, int h) {
  auto rw = vtkSmartPointer<vtkRenderWindow>::New();
  rw->SetOffScreenRendering(1);
  rw->SetShowWindow(false);
  rw->SetSize(w, h);
  return rw;
}

// 闸门：最小离屏渲染冒烟。返回 0=OK，非 0=离屏 GL 起不来（BLOCKED_OFFSCREEN）。
// 流程：离屏窗口 → 加一个 cube actor → Render() → 读回像素，确认非全黑/读得到。
int cmdOffscreenSmoke() {
  std::cout << "[offscreen-smoke] 创建离屏 vtkRenderWindow...\n";
  try {
    auto rw = makeOffscreenWindow(256, 256);

    vtkNew<vtkRenderer> ren;
    ren->SetBackground(0.1, 0.1, 0.2);
    rw->AddRenderer(ren);

    vtkNew<vtkCubeSource> cube;
    cube->SetXLength(1.0);
    cube->SetYLength(1.0);
    cube->SetZLength(1.0);

    vtkNew<vtkPolyDataMapper> mapper;
    mapper->SetInputConnection(cube->GetOutputPort());

    vtkNew<vtkActor> actor;
    actor->SetMapper(mapper);
    actor->GetProperty()->SetColor(1.0, 0.6, 0.2);
    ren->AddActor(actor);
    ren->ResetCamera();

    // Render()：若 GL 上下文创建失败，VTK 会输出错误（多数返回，少数抛）。
    rw->Render();

    // 读回像素验证：取整窗 RGB，确认能读到且非全 0。
    const int* sz = rw->GetSize();
    const int w = sz[0], h = sz[1];
    if (w <= 0 || h <= 0) {
      std::cout << "[offscreen-smoke] FAIL: 窗口尺寸为 0（上下文未建立）\n";
      std::cout << "STATUS=BLOCKED_OFFSCREEN\n";
      return 1;
    }

    auto pixels = vtkSmartPointer<vtkUnsignedCharArray>::New();
    // GetRGBACharPixelData(x0,y0,x1,y1,front,arr)：front=1 读前缓冲。
    const int ok =
        rw->GetRGBACharPixelData(0, 0, w - 1, h - 1, /*front=*/1, pixels);
    if (ok == 0 || pixels->GetNumberOfTuples() == 0) {
      std::cout << "[offscreen-smoke] FAIL: 读不到像素\n";
      std::cout << "STATUS=BLOCKED_OFFSCREEN\n";
      return 1;
    }

    // 统计非背景像素（cube 应渲出橙色，存在像素 R 通道明显高于背景）。
    vtkIdType nonBlack = 0;
    const vtkIdType n = pixels->GetNumberOfTuples();
    for (vtkIdType i = 0; i < n; ++i) {
      const double r = pixels->GetComponent(i, 0);
      const double g = pixels->GetComponent(i, 1);
      const double b = pixels->GetComponent(i, 2);
      if (r > 80 || g > 80 || b > 80) ++nonBlack;
    }

    const char* caps = rw->ReportCapabilities();
    std::cout << "[offscreen-smoke] 读回像素 " << n << " 个，非背景像素 "
              << nonBlack << "\n";
    std::cout << "[offscreen-smoke] GL 能力:\n"
              << (caps ? caps : "(无)") << "\n";

    if (nonBlack == 0) {
      std::cout << "[offscreen-smoke] FAIL: 渲染结果全为背景（actor 未画出）\n";
      std::cout << "STATUS=BLOCKED_OFFSCREEN\n";
      return 1;
    }

    std::cout << "[offscreen-smoke] OK：离屏 GL 可用，可继续真实基准。\n";
    std::cout << "STATUS=OK\n";
    return 0;
  } catch (const std::exception& e) {
    std::cout << "[offscreen-smoke] FAIL: 异常 " << e.what() << "\n";
    std::cout << "STATUS=BLOCKED_OFFSCREEN\n";
    return 1;
  }
}

// 体绘制 fps：每帧绕 azimuth 旋相机再 Render()，避免被驱动优化成空渲染。
double benchVolumeFps(vtkRenderWindow* rw, vtkRenderer* ren, int frames) {
  ren->ResetCamera();
  vtkCamera* cam = ren->GetActiveCamera();
  rw->Render();  // 预热一帧（首帧含上传显存/编译 shader，不计时）
  Stopwatch sw;
  for (int f = 0; f < frames; ++f) {
    cam->Azimuth(360.0 / frames);  // 每帧转一点，扫满一圈
    rw->Render();
  }
  const double ms = sw.elapsedMs();
  return ms > 0.0 ? frames * 1000.0 / ms : 0.0;
}

// 切片扫描 fps：沿 K 轴(深度)逐偏移 reslice 取轴向切面 + 纹理渲染，每帧推进偏移。
double benchSliceFps(vtkRenderWindow* rw, vtkRenderer* ren,
                     vtkImageData* full, vtkLookupTable* lut, int frames) {
  // reslice：固定轴向(XY 平面)，沿 Z 改变 ResliceAxesOrigin 扫过整卷。
  vtkNew<vtkImageReslice> reslice;
  reslice->SetInputData(full);
  reslice->SetOutputDimensionality(2);
  reslice->SetInterpolationModeToLinear();

  vtkNew<vtkImageMapToColors> colorize;
  colorize->SetLookupTable(lut);
  colorize->SetInputConnection(reslice->GetOutputPort());

  vtkNew<vtkImageActor> imgActor;
  imgActor->GetMapper()->SetInputConnection(colorize->GetOutputPort());
  ren->AddViewProp(imgActor);
  ren->ResetCamera();

  double bounds[6];
  full->GetBounds(bounds);
  const double zMin = bounds[4], zMax = bounds[5];
  const double ox = 0.5 * (bounds[0] + bounds[1]);
  const double oy = 0.5 * (bounds[2] + bounds[3]);

  rw->Render();  // 预热
  Stopwatch sw;
  for (int f = 0; f < frames; ++f) {
    const double t = static_cast<double>(f) / std::max(1, frames - 1);
    const double z = zMin + (zMax - zMin) * t;
    reslice->SetResliceAxesOrigin(ox, oy, z);
    reslice->Modified();
    rw->Render();
  }
  const double ms = sw.elapsedMs();
  return ms > 0.0 ? frames * 1000.0 / ms : 0.0;
}

// 由 ColorScale 物理区间建 256 级 VTK LUT（切片纹理着色用，与体绘制色阶同源）。
vtkSmartPointer<vtkLookupTable> makeLut(const geopro::core::ColorScale& cs,
                                        double vmin, double vmax) {
  auto lut = vtkSmartPointer<vtkLookupTable>::New();
  const int n = 256;
  lut->SetNumberOfTableValues(n);
  lut->SetRange(vmin, vmax);
  for (int i = 0; i < n; ++i) {
    const double v = vmin + (vmax - vmin) * i / (n - 1);
    const auto c = cs.colorAt(v);
    lut->SetTableValue(i, c.r / 255.0, c.g / 255.0, c.b / 255.0, 1.0);
  }
  lut->Build();
  return lut;
}

// 简单蓝-白-红色阶（与 test_color_scale 同款最简构造）。
geopro::core::ColorScale makeColorScale(double vmin, double vmax) {
  geopro::core::ColorScale cs;
  const double mid = 0.5 * (vmin + vmax);
  cs.addStop(vmin, geopro::core::Rgba{0, 0, 255, 255});
  cs.addStop(mid, geopro::core::Rgba{255, 255, 255, 255});
  cs.addStop(vmax, geopro::core::Rgba{255, 0, 0, 255});
  return cs;
}

int cmdRenderB(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr << "用法: gpr_poc renderB <storeDir> [--frames 120]\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  const int frames = std::stoi(a.get("frames", "120"));
  std::cout << "[renderB] storeDir=" << dir << " frames=" << frames << "\n";

  // 闸门复检：renderB 前先确认离屏可用（避免在不可渲染机上跑出假数据）。
  std::cout << "[renderB] 离屏闸门复检...\n";
  if (cmdOffscreenSmoke() != 0) {
    std::cout << "[renderB] 闸门失败，中止，不产出 fps。\n";
    return 1;
  }

  // 1) 加载整卷（VTK_SHORT）。
  Stopwatch swLoad;
  geopro::render::WholeVolumeSource src(dir);
  const double loadMs = swLoad.elapsedMs();
  const auto& m = src.meta();

  const std::int64_t voxels =
      static_cast<std::int64_t>(m.nx) * m.ny * m.nz;
  const std::int64_t wholeBytes = voxels * 2;  // VTK_SHORT
  std::cout << "[renderB] 整卷 " << m.nx << "x" << m.ny << "x" << m.nz
            << " 体素=" << voxels << " 字节=" << wholeBytes << " ("
            << wholeBytes / (1024.0 * 1024.0) << " MB)，加载 " << loadMs
            << "ms\n";

  auto images = src.currentImages();
  if (images.empty() || !images.front()) {
    std::cerr << "[renderB] 错误: currentImages 为空\n";
    return 1;
  }
  vtkImageData* shortImg = images.front().Get();

  // 色阶用 meta 的物理区间。
  const double vmin = m.vminPhys, vmax = m.vmaxPhys;
  const geopro::core::ColorScale cs = makeColorScale(vmin, vmax);

  // 2) 体绘制（离屏）。
  auto rw = makeOffscreenWindow(1024, 768);
  vtkNew<vtkRenderer> ren;
  ren->SetBackground(0.0, 0.0, 0.0);
  rw->AddRenderer(ren);

  vtkSmartPointer<vtkVolume> volume =
      geopro::render::buildVoxelI16FromImage(shortImg, m.quant, cs, vmin, vmax);
  ren->AddVolume(volume);

  // 装上捕获式 OutputWindow：拦截体绘制时的 3D 纹理维度错误。
  auto capWin = vtkSmartPointer<CapturingOutputWindow>::New();
  vtkOutputWindow::SetInstance(capWin);

  std::cout << "[renderB] 体绘制基准（" << frames << " 帧旋转相机）...\n";
  const double volFpsRaw = benchVolumeFps(rw, ren, frames);

  const bool textureErr = capWin->textureError();
  vtkOutputWindow::SetInstance(nullptr);  // 还原默认输出窗口

  // 进显存判据：SmartVolumeMapper 实际用的渲染模式（2=GPURenderMode）。
  int renderMode = -1;
  bool lowResResample = false;
  if (auto* svm =
          vtkSmartVolumeMapper::SafeDownCast(volume->GetMapper())) {
    renderMode = svm->GetLastUsedRenderMode();
    // 大体可能触发降质重采样（GPU 显存不足时 SmartVolumeMapper 走低分辨率）。
    lowResResample = (svm->GetInteractiveAdjustSampleDistances() == 0 &&
                      renderMode != vtkSmartVolumeMapper::GPURenderMode);
  }
  const bool onGpu = (renderMode == vtkSmartVolumeMapper::GPURenderMode);

  // 任一维度超过 GL_MAX_3D_TEXTURE_SIZE（本机实测 16384）→ 整卷无法成单张 3D 纹理。
  constexpr int kMax3DTexObserved = 16384;
  const bool dimOversize =
      (m.nx > kMax3DTexObserved || m.ny > kMax3DTexObserved ||
       m.nz > kMax3DTexObserved);
  // 体绘制 fps 是否可信：上传成功（无纹理错误且未超限）才算真实整卷体绘制帧率。
  const bool volFpsValid = !textureErr && !dimOversize;
  const double volFps = volFpsValid ? volFpsRaw : -1.0;

  std::cout << "[renderB] 体绘制 raw_fps=" << volFpsRaw
            << "  渲染模式=" << renderMode << (onGpu ? "(GPU)" : "(非GPU)")
            << "  纹理维度错误=" << (textureErr ? "是" : "否")
            << "  超 16384=" << (dimOversize ? "是" : "否") << "\n";
  if (!volFpsValid) {
    std::cout << "[renderB] 警告: 整卷未能成单张 3D 纹理（X=" << m.nx
              << " > " << kMax3DTexObserved
              << "），体绘制 fps 无意义 → 标 INVALID。\n";
  }

  // 3) 切片扫描（离屏，沿 Z 扫整卷）。
  vtkNew<vtkRenderer> ren2;
  ren2->SetBackground(0.0, 0.0, 0.0);
  auto rw2 = makeOffscreenWindow(1024, 768);
  rw2->AddRenderer(ren2);
  vtkSmartPointer<vtkLookupTable> lut = makeLut(cs, vmin, vmax);

  std::cout << "[renderB] 切片扫描基准（" << frames << " 帧沿 Z 推进）...\n";
  const double sliceFps =
      benchSliceFps(rw2, ren2, src.sliceSource(), lut, frames);
  std::cout << "[renderB] 切片 fps=" << sliceFps << "\n";

  const double peak = Probe::peakMemMB();
  const std::string vram = "N/A";  // VTK 安装未带 GLEW 头，无法直查 NVX 显存

  // 4) 汇总打印。
  const std::string volFpsStr =
      volFpsValid ? std::to_string(volFps) : "INVALID(整卷超 3D 纹理上限)";

  std::cout << "\n=== renderB GPU 指标 ===\n";
  std::cout << "离屏闸门       : OK\n";
  std::cout << "体维度         : " << m.nx << " x " << m.ny << " x " << m.nz
            << "\n";
  std::cout << "体素数         : " << voxels << "\n";
  std::cout << "整卷字节(B)    : " << wholeBytes << "  ("
            << wholeBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "体绘制 fps     : " << volFpsStr << "\n";
  if (!volFpsValid) {
    std::cout << "  (raw_fps=" << volFpsRaw
              << " 为空纹理渲染，X=" << m.nx << " > 16384，不可信)\n";
  }
  std::cout << "切片扫描 fps   : " << sliceFps << "  (2D 纹理，无 3D 上限约束)\n";
  std::cout << "渲染模式       : " << renderMode
            << (onGpu ? " (GPU 路径)" : " (非 GPU)") << "\n";
  std::cout << "整卷进显存     : "
            << (volFpsValid && onGpu ? "是（单张 3D 纹理）"
                                     : "否（超 GL_MAX_3D_TEXTURE_SIZE 16384）")
            << "\n";
  std::cout << "降质重采样     : " << (lowResResample ? "是" : "否") << "\n";
  std::cout << "GPU 显存       : " << vram << "\n";
  std::cout << "进程峰值内存(MB): " << peak << "\n";

  writeMetricLine(
      "renderB,dir=" + dir + ",nx=" + std::to_string(m.nx) +
      ",ny=" + std::to_string(m.ny) + ",nz=" + std::to_string(m.nz) +
      ",voxels=" + std::to_string(voxels) +
      ",wholeB=" + std::to_string(wholeBytes) +
      ",volFps=" + volFpsStr +
      ",volFpsRaw=" + std::to_string(volFpsRaw) +
      ",volFpsValid=" + std::to_string(volFpsValid ? 1 : 0) +
      ",sliceFps=" + std::to_string(sliceFps) +
      ",renderMode=" + std::to_string(renderMode) +
      ",onGpu=" + std::to_string(onGpu ? 1 : 0) +
      ",loadMs=" + std::to_string(loadMs) +
      ",peakMB=" + std::to_string(peak));
  return 0;
}

// ============================================================================
// 核外分块体绘制基准（POC-C，命门探针）
// ============================================================================

// 量化域传函(与 VoxelActor::buildVoxelI16FromImage 同逻辑):颜色对每量化级 qv 用
// q.toPhys(qv) 反查 ColorScale;不透明度 kBlank→0、[qmin,qmax] 线性到 kMaxOpacity。
// MultiBlock 全块共用同一 vtkVolumeProperty(挂在单个 vtkVolume 上)。
vtkSmartPointer<vtkVolumeProperty> makeI16VolumeProperty(
    const geopro::core::Quant& q, const geopro::core::ColorScale& cs,
    double vminPhys, double vmaxPhys) {
  constexpr int kTransferSamples = 64;
  constexpr double kMaxOpacity = 0.15;
  if (vminPhys >= vmaxPhys) vmaxPhys = vminPhys + 1.0;

  const double qminD = static_cast<double>(q.toQ(vminPhys));
  const double qmaxD = static_cast<double>(q.toQ(vmaxPhys));

  vtkNew<vtkColorTransferFunction> color;
  for (int t = 0; t < kTransferSamples; ++t) {
    const double qd = qminD + (qmaxD - qminD) * t / (kTransferSamples - 1);
    const auto qvLevel = static_cast<std::int16_t>(std::lround(qd));
    const double phys = q.toPhys(qvLevel);
    const auto c = cs.colorAt(phys);
    color->AddRGBPoint(qd, c.r / 255.0, c.g / 255.0, c.b / 255.0);
  }

  vtkNew<vtkPiecewiseFunction> opacity;
  opacity->AddPoint(
      static_cast<double>(geopro::core::ScalarVolumeI16::kBlank), 0.0);
  opacity->AddPoint(qminD, 0.0);
  opacity->AddPoint(qmaxD, kMaxOpacity);

  auto prop = vtkSmartPointer<vtkVolumeProperty>::New();
  prop->SetColor(color);
  prop->SetScalarOpacity(opacity);
  prop->SetInterpolationTypeToLinear();
  prop->ShadeOff();
  return prop;
}

// 由当前工作集图像组装 vtkMultiBlockDataSet(每块一个 vtkImageData)。
vtkSmartPointer<vtkMultiBlockDataSet> makeMultiBlock(
    const std::vector<vtkSmartPointer<vtkImageData>>& imgs) {
  auto mb = vtkSmartPointer<vtkMultiBlockDataSet>::New();
  mb->SetNumberOfBlocks(static_cast<unsigned int>(imgs.size()));
  for (unsigned int i = 0; i < imgs.size(); ++i) {
    mb->SetBlock(i, imgs[i].Get());
  }
  return mb;
}

int cmdRenderC(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr << "用法: gpr_poc renderC <storeDir> [--budget 64] [--frames 120]\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  const std::size_t budget =
      static_cast<std::size_t>(std::stoul(a.get("budget", "64")));
  const int frames = std::stoi(a.get("frames", "120"));
  std::cout << "[renderC] storeDir=" << dir << " budget=" << budget
            << " frames=" << frames << "\n";

  // 闸门复检:不可渲染机不产假 fps。
  std::cout << "[renderC] 离屏闸门复检...\n";
  if (cmdOffscreenSmoke() != 0) {
    std::cout << "[renderC] 闸门失败,中止,不产出 fps。\n";
    return 1;
  }

  // 1) 核外源(读 meta + 建 pager,不载整卷)。
  Stopwatch swLoad;
  geopro::render::OutOfCoreSource src(dir, budget);
  const double loadMs = swLoad.elapsedMs();
  const auto& m = src.meta();
  const std::int64_t voxels =
      static_cast<std::int64_t>(m.nx) * m.ny * m.nz;

  const int winW = 1024, winH = 768;
  src.setAspect(static_cast<double>(winW) / winH);

  std::cout << "[renderC] 体 " << m.nx << "x" << m.ny << "x" << m.nz
            << " 体素=" << voxels << " (整卷 X=" << m.nx
            << " > 16384 → renderB INVALID)，源构造 " << loadMs << "ms\n";

  // 色阶用 meta 物理区间。
  const double vmin = m.vminPhys, vmax = m.vmaxPhys;
  const geopro::core::ColorScale cs = makeColorScale(vmin, vmax);
  vtkSmartPointer<vtkVolumeProperty> prop =
      makeI16VolumeProperty(m.quant, cs, vmin, vmax);

  // 2) 离屏 + MultiBlock 体绘制。
  auto rw = makeOffscreenWindow(winW, winH);
  vtkNew<vtkRenderer> ren;
  ren->SetBackground(0.0, 0.0, 0.0);
  rw->AddRenderer(ren);

  vtkNew<vtkMultiBlockVolumeMapper> mapper;
  mapper->SetRequestedRenderMode(vtkSmartVolumeMapper::GPURenderMode);

  auto volume = vtkSmartPointer<vtkVolume>::New();
  volume->SetMapper(mapper);
  volume->SetProperty(prop);
  ren->AddVolume(volume);

  // 装捕获式 OutputWindow:拦截每块上传时的 3D 纹理维度错误(应无,因块 ≤64³)。
  auto capWin = vtkSmartPointer<CapturingOutputWindow>::New();
  vtkOutputWindow::SetInstance(capWin);

  // 相机:先以全体定向(看整卷),首帧 update 选出工作集后再 ResetCamera 到
  // 实际驻留块的 mapper 包围盒(budget<视野总块时工作集只覆盖体的一部分,框住它
  // 才能确证核外体绘制真渲出;这是 budget 受限下的诚实测法,报告说明)。
  ren->ResetCamera(m.origin[0], m.origin[0] + m.nx * m.spacing[0],
                   m.origin[1], m.origin[1] + m.ny * m.spacing[1],
                   m.origin[2], m.origin[2] + m.nz * m.spacing[2]);
  vtkCamera* cam = ren->GetActiveCamera();

  auto refreshBlocks = [&]() {
    src.update(cam);
    auto imgs = src.currentImages();
    auto mb = makeMultiBlock(imgs);
    mapper->SetInputDataObject(mb);
    mapper->Update();
    return imgs.size();
  };

  const std::size_t warmBlocks = refreshBlocks();
  // 用工作集(mapper)实际包围盒重置相机,框住驻留块。
  {
    double b[6];
    mapper->GetBounds(b);
    if (b[0] <= b[1]) {
      ren->ResetCamera(b);
    } else {
      ren->ResetCamera();
    }
  }
  rw->Render();  // 预热(上传显存 + 编译 shader,不计时)

  {
    double b[6];
    mapper->GetBounds(b);
    std::cout << "[renderC] 工作集包围盒 x[" << b[0] << "," << b[1] << "] y["
              << b[2] << "," << b[3] << "] z[" << b[4] << "," << b[5] << "]\n";
  }

  std::cout << "[renderC] 预热:level=" << src.lastLevel()
            << " 视野块=" << src.lastVisibleCount() << "/"
            << src.lastLevelBrickTotal()
            << " 驻留=" << src.residentCount() << " 渲染块=" << warmBlocks
            << "\n";

  std::size_t maxResident = src.residentCount();
  std::size_t sumBlocks = 0;

  // 3a) 静态工作集体绘制 fps:工作集固定(不每帧换块),只旋相机 + Render。
  // 隔离"纯 GPU MultiBlock 体绘制"成本(剔除分块换页/解压/重建 mapper 开销),
  // 直接对照 renderB 整卷 fps,回答未知 #6(真实体绘制 fps)。
  std::cout << "[renderC] 静态工作集体绘制基准(" << frames << " 帧旋相机)...\n";
  Stopwatch swStatic;
  for (int f = 0; f < frames; ++f) {
    cam->Azimuth(360.0 / frames);
    rw->Render();  // 工作集不变,仅旋转
  }
  const double staticMs = swStatic.elapsedMs();
  const double staticFps = staticMs > 0 ? frames * 1000.0 / staticMs : 0.0;
  std::cout << "[renderC] 静态工作集 fps=" << staticFps << "\n";

  // 3b) 动态换页体绘制 fps:每帧 update(cam)(重选 LOD/视野块,含 qUncompress 解压
  // 换入的块 + 重建 MultiBlock)+ Render。回答未知 #4(热路径解压是否拖垮 fps)
  // 与 #5(内存恒定)。同时累计 update 耗时占比。
  std::cout << "[renderC] 动态换页体绘制基准(" << frames << " 帧旋相机)...\n";
  double updateMsTotal = 0.0;
  Stopwatch swDyn;
  for (int f = 0; f < frames; ++f) {
    cam->Azimuth(360.0 / frames);
    Stopwatch swU;
    const std::size_t blocks = refreshBlocks();  // update + 重建 MultiBlock
    updateMsTotal += swU.elapsedMs();
    sumBlocks += blocks;
    maxResident = std::max(maxResident, src.residentCount());
    rw->Render();
  }
  const double dynMs = swDyn.elapsedMs();
  const double dynFps = dynMs > 0 ? frames * 1000.0 / dynMs : 0.0;
  const double rawFps = dynFps;  // 主报告口径:含换页的真实交互 fps
  std::cout << "[renderC] 动态换页 fps=" << dynFps
            << "  (其中 update/换页/重建 平均 " << (updateMsTotal / frames)
            << " ms/帧)\n";

  const bool textureErr = capWin->textureError();
  vtkOutputWindow::SetInstance(nullptr);

  // 4) 正确性判据:渲出非空像素(非全背景)。
  auto pixels = vtkSmartPointer<vtkUnsignedCharArray>::New();
  rw->GetRGBACharPixelData(0, 0, winW - 1, winH - 1, /*front=*/1, pixels);
  vtkIdType nonBlack = 0;
  const vtkIdType npx = pixels->GetNumberOfTuples();
  for (vtkIdType i = 0; i < npx; ++i) {
    if (pixels->GetComponent(i, 0) > 10 || pixels->GetComponent(i, 1) > 10 ||
        pixels->GetComponent(i, 2) > 10) {
      ++nonBlack;
    }
  }
  const bool renderedNonEmpty = (nonBlack > 0);

  // 渲染模式(MultiBlock 内部每块一个 SmartVolumeMapper;此处取一块代表性查询)。
  // MultiBlock 不直接暴露 LastUsedRenderMode,故以纹理无错 + 非空像素为体绘制真出证据。
  const bool volFpsValid = !textureErr && renderedNonEmpty;

  const double peak = Probe::peakMemMB();
  const double avgBlocks =
      frames > 0 ? static_cast<double>(sumBlocks) / frames : 0.0;

  std::cout << "\n=== renderC 核外体绘制指标 ===\n";
  std::cout << "离屏闸门         : OK\n";
  std::cout << "体维度           : " << m.nx << " x " << m.ny << " x " << m.nz
            << "  (整卷 X 超 16384,renderB=INVALID)\n";
  std::cout << "体素数           : " << voxels << "\n";
  std::cout << "budget(块)       : " << budget << "\n";
  std::cout << "峰值驻留(块)     : " << maxResident
            << (maxResident <= budget ? "  (≤budget,内存恒定 OK)"
                                      : "  (!! 超 budget)")
            << "\n";
  std::cout << "末帧 level       : " << src.lastLevel() << "\n";
  std::cout << "末帧视野块/总块  : " << src.lastVisibleCount() << " / "
            << src.lastLevelBrickTotal() << "\n";
  std::cout << "平均渲染块/帧    : " << avgBlocks << "\n";
  std::cout << "纹理维度错误     : " << (textureErr ? "是(!!)" : "否") << "\n";
  std::cout << "渲出非空像素     : " << (renderedNonEmpty ? "是" : "否(!!)")
            << "  (非背景像素=" << nonBlack << ")\n";
  std::cout << "静态工作集 fps   : "
            << (volFpsValid ? std::to_string(staticFps)
                            : std::string("INVALID(纹理错或空渲染)"))
            << "  (纯 GPU MultiBlock 体绘制)\n";
  std::cout << "动态换页 fps     : "
            << (volFpsValid ? std::to_string(dynFps)
                            : std::string("INVALID(纹理错或空渲染)"))
            << "  (含每帧 update/解压/重建 mapper)\n";
  std::cout << "  换页均耗时/帧  : " << (updateMsTotal / frames) << " ms\n";
  std::cout << "进程峰值内存(MB) : " << peak << "\n";
  std::cout << "源构造耗时(ms)   : " << loadMs << "\n";
  std::cout << "对照 renderB     : 整卷 INVALID(超 3D 纹理上限);renderC "
            << (volFpsValid ? "真渲出 ✔" : "未渲出 ✘") << "\n";

  writeMetricLine(
      "renderC,dir=" + dir + ",nx=" + std::to_string(m.nx) +
      ",ny=" + std::to_string(m.ny) + ",nz=" + std::to_string(m.nz) +
      ",voxels=" + std::to_string(voxels) +
      ",budget=" + std::to_string(budget) +
      ",maxResident=" + std::to_string(maxResident) +
      ",lastLevel=" + std::to_string(src.lastLevel()) +
      ",lastVisible=" + std::to_string(src.lastVisibleCount()) +
      ",lastLevelTotal=" + std::to_string(src.lastLevelBrickTotal()) +
      ",avgBlocks=" + std::to_string(avgBlocks) +
      ",textureErr=" + std::to_string(textureErr ? 1 : 0) +
      ",nonBlack=" + std::to_string(nonBlack) +
      ",volFpsValid=" + std::to_string(volFpsValid ? 1 : 0) +
      ",staticFps=" + (volFpsValid ? std::to_string(staticFps) : "INVALID") +
      ",dynFps=" + (volFpsValid ? std::to_string(dynFps) : "INVALID") +
      ",updateMsPerFrame=" + std::to_string(updateMsTotal / frames) +
      ",rawFps=" + std::to_string(rawFps) +
      ",loadMs=" + std::to_string(loadMs) +
      ",peakMB=" + std::to_string(peak));
  return volFpsValid ? 0 : 1;
}

// ============================================================================
// 单 mapper SetPartitions 整卷体绘制基准（POC-C-partitioned,去风险探针）
// ============================================================================
//
// 验"对的架构":整卷喂【单个】vtkGPUVolumeRayCastMapper(其 OpenGL 实现 =
// vtkOpenGLGPUVolumeRayCastMapper),用 SetPartitions(ceil(nx/16384),...) 让同一
// mapper 内部把体沿轴分区上传(每区 ≤16384 绕过 GL_MAX_3D_TEXTURE_SIZE),一次
// ray cast。对照 9c 整卷单 SmartVolumeMapper(INVALID,纹理墙) 与 12 MultiBlock
// (每块一 mapper,9.5 静态/1.45 换页)。
//
// 双闸(同 9c,绝不把空纹理假帧率当性能):
//   ① CapturingOutputWindow 捕获 3D 纹理维度错误;
//   ② 真实回读像素,统计非背景像素 → 非空才算真渲出。
int cmdRenderCPartitioned(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr
        << "用法: gpr_poc renderC-partitioned <storeDir> [--frames 120]\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  const int frames = std::stoi(a.get("frames", "120"));
  std::cout << "[renderC-partitioned] storeDir=" << dir << " frames=" << frames
            << "\n";

  // 闸门复检:不可渲染机不产假 fps。
  std::cout << "[renderC-partitioned] 离屏闸门复检...\n";
  if (cmdOffscreenSmoke() != 0) {
    std::cout << "[renderC-partitioned] 闸门失败,中止,不产出 fps。\n";
    return 1;
  }

  // 1) WholeVolumeSource 重组整卷 VTK_SHORT image(常驻内存,约 400MB)。
  Stopwatch swLoad;
  geopro::render::WholeVolumeSource src(dir);
  const double loadMs = swLoad.elapsedMs();
  const auto& m = src.meta();

  const std::int64_t voxels =
      static_cast<std::int64_t>(m.nx) * m.ny * m.nz;
  const std::int64_t wholeBytes = voxels * 2;  // VTK_SHORT
  std::cout << "[renderC-partitioned] 整卷 " << m.nx << "x" << m.ny << "x"
            << m.nz << " 体素=" << voxels << " 字节=" << wholeBytes << " ("
            << wholeBytes / (1024.0 * 1024.0) << " MB),加载 " << loadMs
            << "ms\n";

  auto images = src.currentImages();
  if (images.empty() || !images.front()) {
    std::cerr << "[renderC-partitioned] 错误: currentImages 为空\n";
    return 1;
  }
  vtkImageData* shortImg = images.front().Get();

  // 2) 分区数:任一轴 > 16384 → ceil(dim/16384) 个分区,其余轴 1。
  constexpr int kMax3DTex = 16384;
  auto partCount = [](int dim) {
    return static_cast<unsigned short>((dim + kMax3DTex - 1) / kMax3DTex);
  };
  const unsigned short px = partCount(m.nx);
  const unsigned short py = partCount(m.ny);
  const unsigned short pz = partCount(m.nz);
  std::cout << "[renderC-partitioned] SetPartitions(" << px << "," << py << ","
            << pz << ")  每区上限 ≤" << kMax3DTex << " (沿线 " << m.nx << "/"
            << px << "=" << (m.nx + px - 1) / px << ")\n";

  // 3) 量化域传函(复用现有 makeI16VolumeProperty:qmin/qmax + kBlank 透明)。
  const double vmin = m.vminPhys, vmax = m.vmaxPhys;
  const geopro::core::ColorScale cs = makeColorScale(vmin, vmax);
  vtkSmartPointer<vtkVolumeProperty> prop =
      makeI16VolumeProperty(m.quant, cs, vmin, vmax);

  // 4) 离屏 + 单个 GPU ray cast mapper + SetPartitions。
  const int winW = 1024, winH = 768;
  auto rw = makeOffscreenWindow(winW, winH);
  vtkNew<vtkRenderer> ren;
  ren->SetBackground(0.0, 0.0, 0.0);
  rw->AddRenderer(ren);

  // vtkGPUVolumeRayCastMapper 抽象基类无 SetPartitions(在 OpenGL 实现上);
  // 直接建 OpenGL 具体类(工厂默认产物同此),喂【整卷单 image】不预切块。
  vtkNew<vtkOpenGLGPUVolumeRayCastMapper> mapper;
  mapper->SetInputData(shortImg);
  mapper->SetPartitions(px, py, pz);

  auto volume = vtkSmartPointer<vtkVolume>::New();
  volume->SetMapper(mapper);
  volume->SetProperty(prop);
  ren->AddVolume(volume);

  // 装捕获式 OutputWindow:拦截分区上传时的 3D 纹理维度错误。
  auto capWin = vtkSmartPointer<CapturingOutputWindow>::New();
  vtkOutputWindow::SetInstance(capWin);

  // 相机:用 mapper 实际包围盒定向(整卷,非工作集);体极扁长(44476:29:162),
  // ResetCamera 全体后再倾斜抬高视角,让薄维度可见(否则边缘视角近乎不可见)。
  {
    double b[6];
    mapper->GetBounds(b);
    if (b[0] <= b[1]) {
      ren->ResetCamera(b);
    } else {
      ren->ResetCamera();
    }
  }
  vtkCamera* cam = ren->GetActiveCamera();
  cam->Elevation(30.0);   // 抬高,避免纯边缘视角看不到薄板
  cam->Azimuth(30.0);
  ren->ResetCameraClippingRange();

  // 每帧旋相机 + Render 测 fps;同时多帧采样非背景像素取最大值
  // (区分"真渲不出"与"末帧恰好边缘视角空"——后者只是采样时机)。
  auto countNonBlack = [&]() -> vtkIdType {
    auto px = vtkSmartPointer<vtkUnsignedCharArray>::New();
    rw->GetRGBACharPixelData(0, 0, winW - 1, winH - 1, /*front=*/1, px);
    vtkIdType nb = 0;
    const vtkIdType np = px->GetNumberOfTuples();
    for (vtkIdType i = 0; i < np; ++i) {
      if (px->GetComponent(i, 0) > 10 || px->GetComponent(i, 1) > 10 ||
          px->GetComponent(i, 2) > 10) {
        ++nb;
      }
    }
    return nb;
  };

  std::cout << "[renderC-partitioned] 单 mapper 整卷体绘制基准(" << frames
            << " 帧旋相机)...\n";
  rw->Render();  // 预热(分区上传 + 编译 shader,不计时)
  vtkIdType maxNonBlack = countNonBlack();
  const int sampleEvery = std::max(1, frames / 8);
  Stopwatch swBench;
  for (int f = 0; f < frames; ++f) {
    cam->Azimuth(360.0 / frames);
    rw->Render();
    if (f % sampleEvery == 0) {
      maxNonBlack = std::max(maxNonBlack, countNonBlack());
    }
  }
  const double benchMs = swBench.elapsedMs();
  const double volFpsRaw =
      benchMs > 0.0 ? frames * 1000.0 / benchMs : 0.0;

  const bool textureErr = capWin->textureError();
  vtkOutputWindow::SetInstance(nullptr);

  // 5) 正确性判据:整个旋转扫描中的最大非背景像素(非空才算真渲出)。
  const vtkIdType nonBlack = maxNonBlack;
  const bool renderedNonEmpty = (nonBlack > 0);

  // 双闸:无纹理错 + 非空像素 → fps 可信。
  const bool volFpsValid = !textureErr && renderedNonEmpty;
  const double volFps = volFpsValid ? volFpsRaw : -1.0;
  const double peak = Probe::peakMemMB();
  const bool interactive = volFpsValid && volFps >= 15.0;

  const std::string volFpsStr =
      volFpsValid ? std::to_string(volFps)
                  : std::string("INVALID(纹理错或空渲染)");

  std::cout << "\n=== renderC-partitioned 单 mapper SetPartitions 指标 ===\n";
  std::cout << "离屏闸门         : OK\n";
  std::cout << "体维度           : " << m.nx << " x " << m.ny << " x " << m.nz
            << "\n";
  std::cout << "体素数           : " << voxels << "\n";
  std::cout << "整卷字节(B)      : " << wholeBytes << "  ("
            << wholeBytes / (1024.0 * 1024.0) << " MB)\n";
  std::cout << "分区数(px,py,pz) : " << px << "," << py << "," << pz << "\n";
  std::cout << "纹理维度错误     : " << (textureErr ? "是(!!)" : "否") << "\n";
  std::cout << "渲出非空像素     : " << (renderedNonEmpty ? "是" : "否(!!)")
            << "  (非背景像素=" << nonBlack << ")\n";
  std::cout << "体绘制 fps       : " << volFpsStr << "\n";
  if (!volFpsValid) {
    std::cout << "  (raw_fps=" << volFpsRaw << " 不可信)\n";
  }
  std::cout << "达交互级(≥15fps) : "
            << (interactive ? "是 ✔" : "否 ✘") << "\n";
  std::cout << "进程峰值内存(MB) : " << peak << "\n";
  std::cout << "源构造耗时(ms)   : " << loadMs << "\n";
  std::cout << "对照 renderB     : 整卷单 SmartVolumeMapper=INVALID(纹理墙);"
               "renderC MultiBlock=9.5 静态/1.45 换页;本探针="
            << (volFpsValid ? volFpsStr + "fps" : "INVALID") << "\n";

  writeMetricLine(
      "renderC-partitioned,dir=" + dir + ",nx=" + std::to_string(m.nx) +
      ",ny=" + std::to_string(m.ny) + ",nz=" + std::to_string(m.nz) +
      ",voxels=" + std::to_string(voxels) +
      ",wholeB=" + std::to_string(wholeBytes) +
      ",px=" + std::to_string(px) + ",py=" + std::to_string(py) +
      ",pz=" + std::to_string(pz) +
      ",textureErr=" + std::to_string(textureErr ? 1 : 0) +
      ",nonBlack=" + std::to_string(nonBlack) +
      ",volFpsValid=" + std::to_string(volFpsValid ? 1 : 0) +
      ",volFps=" + volFpsStr + ",volFpsRaw=" + std::to_string(volFpsRaw) +
      ",interactive=" + std::to_string(interactive ? 1 : 0) +
      ",loadMs=" + std::to_string(loadMs) + ",peakMB=" + std::to_string(peak));

  // 写报告文件(覆盖式,含对照表)。
  {
    const fs::path repo =
        fs::path("docs") / "superpowers" / "plans" / "poc-results-C.md";
    fs::create_directories(repo.parent_path());
    std::ofstream rf(repo.string());
    if (rf) {
      rf << "# POC-C 单 mapper SetPartitions 整卷体绘制探针结果\n\n";
      rf << "## 体\n";
      rf << "- 维度: " << m.nx << " x " << m.ny << " x " << m.nz << " (体素 "
         << voxels << ")\n";
      rf << "- 整卷字节: " << wholeBytes << " B ("
         << wholeBytes / (1024.0 * 1024.0) << " MB, VTK_SHORT)\n";
      rf << "- store: " << dir << "\n\n";
      rf << "## 单 mapper SetPartitions\n";
      rf << "- mapper: vtkOpenGLGPUVolumeRayCastMapper (整卷单 image,不预切块)\n";
      rf << "- 分区数: SetPartitions(" << px << ", " << py << ", " << pz
         << ")  每区上限 ≤" << kMax3DTex << "\n";
      rf << "- 纹理维度错误: " << (textureErr ? "是" : "否") << "\n";
      rf << "- 渲出非空像素: " << (renderedNonEmpty ? "是" : "否") << " (非背景像素 "
         << nonBlack << ")\n";
      rf << "- 体绘制 fps: " << volFpsStr << "\n";
      rf << "- 达交互级(≥15fps): " << (interactive ? "是" : "否") << "\n";
      rf << "- 进程峰值内存: " << peak << " MB\n";
      rf << "- 源构造耗时: " << loadMs << " ms\n\n";
      rf << "## 对照表\n\n";
      rf << "| 路径 | 是否渲出 | fps |\n";
      rf << "|---|---|---|\n";
      rf << "| renderB 整卷单 SmartVolumeMapper | INVALID(纹理墙) | — |\n";
      rf << "| renderC MultiBlock(每块一 mapper) | 渲出 | 9.5 静态/1.45 换页 |\n";
      rf << "| renderC-partitioned 单 mapper SetPartitions | "
         << (volFpsValid ? "渲出" : "未渲出") << " | "
         << (volFpsValid ? volFpsStr : std::string("INVALID")) << " |\n\n";
      rf << "## 判据结论\n";
      if (volFpsValid && interactive) {
        rf << "单 mapper SetPartitions 整卷体绘制【真渲出且达交互级】(" << volFps
           << " fps ≥15)。C production 路线钉死可行。\n";
      } else if (volFpsValid) {
        rf << "单 mapper SetPartitions 整卷体绘制【真渲出但未达交互级】(" << volFps
           << " fps <15)。VTK 这条路天花板暴露,需评估 OpenVDS/自建 GL。\n";
      } else {
        rf << "单 mapper SetPartitions 整卷体绘制【未真渲出】(纹理错="
           << (textureErr ? "是" : "否") << ",非空像素="
           << (renderedNonEmpty ? "是" : "否")
           << ")。SetPartitions 未能绕过纹理墙,如实记录。\n";
      }
      std::cout << "[renderC-partitioned] 报告写入 " << repo.string() << "\n";
    }
  }

  return volFpsValid ? 0 : 1;
}

// ============================================================================
// LOD-fps 探针（POC-C 最后一根链子，Task 12c）
// ============================================================================
//
// 12b 已证整卷全分辨率 ray cast(2.08 亿体素)~10fps 是硬上限，fps 杠杆只有 LOD
// （渲更少体素）。本探针在【真实金字塔 store】上验四件事，全离屏、双闸防假帧率：
//   (a) 粗层概览 fps：level2 整卷（单轴 <16384 → 单 SmartVolumeMapper）。
//   (b) 全分辨率局部 fps：level0 一段 brick 列（沿线局部）。
//   (c) LOD 切换动态过渡：相机从远观(level2)逐步拉近到近观局部(level0)，跨越
//       LOD 切换那一下逐帧记帧耗时，标切换帧尖峰/stall。
//   (d) 截图：lod-overview.png / lod-fullres-local.png / lod-transition-mid.png。
//
// 双闸（同 9c，绝不把空纹理假帧率当性能）：
//   ① CapturingOutputWindow 捕获 3D 纹理维度错误；
//   ② 真实回读像素，统计非背景像素 → 非空才算真渲出。

// 把金字塔某 level 重组成整卷 VTK_SHORT vtkImageData（逻辑同 WholeVolumeSource，
// 但按 level 维度 + spacing×2^level，使物理范围与 level0 一致）。
vtkSmartPointer<vtkImageData> buildLevelImage(
    const geopro::data::ChunkedVolumeStore& store, int level,
    const geopro::data::StoreMeta& m) {
  int nx = 0, ny = 0, nz = 0;
  store.dims(level, nx, ny, nz);
  const int brick = m.brick;
  const double sc = static_cast<double>(1 << level);  // 2^level

  auto img = vtkSmartPointer<vtkImageData>::New();
  img->SetDimensions(nx, ny, nz);
  img->SetOrigin(m.origin[0], m.origin[1], m.origin[2]);
  img->SetSpacing(m.spacing[0] * sc, m.spacing[1] * sc, m.spacing[2] * sc);

  vtkNew<vtkShortArray> arr;
  arr->SetName("v");
  arr->SetNumberOfTuples(static_cast<vtkIdType>(nx) * ny * nz);

  for (int bz = 0; bz < store.bricksZ(level); ++bz) {
    for (int by = 0; by < store.bricksY(level); ++by) {
      for (int bx = 0; bx < store.bricksX(level); ++bx) {
        const std::vector<std::int16_t> raw = store.readBrick(level, bx, by, bz);
        const int i0 = bx * brick, j0 = by * brick, k0 = bz * brick;
        const int bw = (nx - i0 < brick) ? (nx - i0) : brick;
        const int bh = (ny - j0 < brick) ? (ny - j0) : brick;
        const int bd = (nz - k0 < brick) ? (nz - k0) : brick;
        std::size_t w = 0;
        for (int kk = 0; kk < bd; ++kk) {
          const vtkIdType gk = static_cast<vtkIdType>(k0 + kk);
          for (int jj = 0; jj < bh; ++jj) {
            const vtkIdType gj = static_cast<vtkIdType>(j0 + jj);
            vtkIdType id = (gk * ny + gj) * nx + i0;
            for (int ii = 0; ii < bw; ++ii) arr->SetValue(id++, raw[w++]);
          }
        }
      }
    }
  }
  img->GetPointData()->SetScalars(arr);
  return img;
}

// 取 level0 一段 brick 列 [bx0, bx0+bxCount) × 全 Y × 全 Z 重组成局部整卷
// VTK_SHORT image（X 维 = bxCount*brick ≤ 几百，远 <16384，单 3D 纹理）。
// Origin 沿 X 偏移到该段起点，spacing 用 level0 原值。
vtkSmartPointer<vtkImageData> buildLocalLevel0Image(
    const geopro::data::ChunkedVolumeStore& store,
    const geopro::data::StoreMeta& m, int bx0, int bxCount) {
  const int brick = m.brick;
  const int nx0 = m.nx, ny0 = m.ny, nz0 = m.nz;
  const int totBx = store.bricksX(0);
  bx0 = std::max(0, std::min(bx0, totBx - 1));
  bxCount = std::max(1, std::min(bxCount, totBx - bx0));

  const int i0Global = bx0 * brick;
  const int localNx = std::min(bxCount * brick, nx0 - i0Global);

  auto img = vtkSmartPointer<vtkImageData>::New();
  img->SetDimensions(localNx, ny0, nz0);
  img->SetOrigin(m.origin[0] + i0Global * m.spacing[0], m.origin[1],
                 m.origin[2]);
  img->SetSpacing(m.spacing[0], m.spacing[1], m.spacing[2]);

  vtkNew<vtkShortArray> arr;
  arr->SetName("v");
  arr->SetNumberOfTuples(static_cast<vtkIdType>(localNx) * ny0 * nz0);

  for (int bz = 0; bz < store.bricksZ(0); ++bz) {
    for (int by = 0; by < store.bricksY(0); ++by) {
      for (int bx = bx0; bx < bx0 + bxCount; ++bx) {
        const std::vector<std::int16_t> raw = store.readBrick(0, bx, by, bz);
        const int gi0 = bx * brick, j0 = by * brick, k0 = bz * brick;
        const int li0 = gi0 - i0Global;  // 局部 X 起点
        const int bw = (nx0 - gi0 < brick) ? (nx0 - gi0) : brick;
        const int bh = (ny0 - j0 < brick) ? (ny0 - j0) : brick;
        const int bd = (nz0 - k0 < brick) ? (nz0 - k0) : brick;
        std::size_t w = 0;
        for (int kk = 0; kk < bd; ++kk) {
          const vtkIdType gk = static_cast<vtkIdType>(k0 + kk);
          for (int jj = 0; jj < bh; ++jj) {
            const vtkIdType gj = static_cast<vtkIdType>(j0 + jj);
            vtkIdType id = (gk * ny0 + gj) * localNx + li0;
            for (int ii = 0; ii < bw; ++ii) arr->SetValue(id++, raw[w++]);
          }
        }
      }
    }
  }
  img->GetPointData()->SetScalars(arr);
  return img;
}

// 统计当前窗口前缓冲非背景像素（>10 任一通道）。
vtkIdType countNonBlackPixels(vtkRenderWindow* rw, int w, int h) {
  auto px = vtkSmartPointer<vtkUnsignedCharArray>::New();
  rw->GetRGBACharPixelData(0, 0, w - 1, h - 1, /*front=*/1, px);
  vtkIdType nb = 0;
  const vtkIdType np = px->GetNumberOfTuples();
  for (vtkIdType i = 0; i < np; ++i) {
    if (px->GetComponent(i, 0) > 10 || px->GetComponent(i, 1) > 10 ||
        px->GetComponent(i, 2) > 10) {
      ++nb;
    }
  }
  return nb;
}

// 离屏窗口截图 → PNG。
void savePng(vtkRenderWindow* rw, const std::string& path) {
  rw->Render();
  vtkNew<vtkWindowToImageFilter> w2i;
  w2i->SetInput(rw);
  w2i->SetInputBufferTypeToRGB();
  w2i->ReadFrontBufferOff();
  w2i->Update();
  vtkNew<vtkPNGWriter> writer;
  writer->SetFileName(path.c_str());
  writer->SetInputConnection(w2i->GetOutputPort());
  writer->Write();
}

int cmdRenderLOD(int argc, char** argv) {
  const Args a = parseArgs(argc, argv, 2);
  if (a.positional.empty()) {
    std::cerr << "用法: gpr_poc renderLOD <storeDir> [--frames 120]\n";
    return 2;
  }
  const std::string dir = a.positional[0];
  const int frames = std::stoi(a.get("frames", "120"));
  std::cout << "[renderLOD] storeDir=" << dir << " frames=" << frames << "\n";

  // 闸门复检：不可渲染机不产假 fps。
  std::cout << "[renderLOD] 离屏闸门复检...\n";
  if (cmdOffscreenSmoke() != 0) {
    std::cout << "[renderLOD] 闸门失败，中止，不产出 fps。\n";
    return 1;
  }

  geopro::data::ChunkedVolumeStore store(dir);
  const geopro::data::StoreMeta& m = store.meta();
  const int totLevels = store.levels();
  std::cout << "[renderLOD] level0=" << m.nx << "x" << m.ny << "x" << m.nz
            << " 总层数=" << totLevels << "\n";
  if (totLevels < 3) {
    std::cout << "[renderLOD] 警告: 金字塔层数 <3（需 build --levels 3）。\n";
  }

  const double vmin = m.vminPhys, vmax = m.vmaxPhys;
  const geopro::core::ColorScale cs = makeColorScale(vmin, vmax);

  const fs::path shotDir =
      fs::path("docs") / "superpowers" / "plans" / "poc-lod-shots";
  fs::create_directories(shotDir);

  const int winW = 1024, winH = 768;

  // 共用一个捕获式 OutputWindow，贯穿三段渲染。
  auto capWin = vtkSmartPointer<CapturingOutputWindow>::New();
  vtkOutputWindow::SetInstance(capWin);

  // ---- (a) 粗层概览 fps：level2 整卷 ----
  const int ovLevel = std::min(2, totLevels - 1);
  std::cout << "[renderLOD] (a) 建 level" << ovLevel << " 整卷 image...\n";
  vtkSmartPointer<vtkImageData> ovImg = buildLevelImage(store, ovLevel, m);
  int ovNx, ovNy, ovNz;
  store.dims(ovLevel, ovNx, ovNy, ovNz);

  auto rwOv = makeOffscreenWindow(winW, winH);
  vtkNew<vtkRenderer> renOv;
  renOv->SetBackground(0.0, 0.0, 0.0);
  rwOv->AddRenderer(renOv);
  vtkSmartPointer<vtkVolume> ovVol =
      geopro::render::buildVoxelI16FromImage(ovImg.Get(), m.quant, cs, vmin,
                                             vmax);
  renOv->AddVolume(ovVol);
  // 先测 fps（benchVolumeFps 内部会 ResetCamera + 旋满一圈）。
  const double ovFps = benchVolumeFps(rwOv.Get(), renOv, frames);
  // 截图前重设一个利于人眼的取景：整线物理纵横比极扁(~2200m×1.5m×8m)，俯视角
  // 看宽面才能呈现整条带(而非边缘线)。
  renOv->ResetCamera();
  renOv->GetActiveCamera()->Elevation(55.0);
  renOv->GetActiveCamera()->Azimuth(20.0);
  renOv->ResetCameraClippingRange();
  rwOv->Render();
  const vtkIdType ovNonBlack = countNonBlackPixels(rwOv.Get(), winW, winH);
  savePng(rwOv.Get(), (shotDir / "lod-overview.png").string());
  std::cout << "[renderLOD] (a) 概览 fps=" << ovFps << " 非空像素=" << ovNonBlack
            << " (level" << ovLevel << " " << ovNx << "x" << ovNy << "x" << ovNz
            << ")\n";

  // ---- (b) 全分辨率局部 fps：level0 一段 brick 列 ----
  const int totBx = store.bricksX(0);
  const int localBx = std::min(4, totBx);            // 4 brick 列 ≈ 256 体素宽
  const int bx0 = std::max(0, totBx / 2 - localBx / 2);  // 取沿线中段
  std::cout << "[renderLOD] (b) 建 level0 局部 image (brick列 [" << bx0 << ","
            << (bx0 + localBx) << ") / " << totBx << ")...\n";
  vtkSmartPointer<vtkImageData> locImg =
      buildLocalLevel0Image(store, m, bx0, localBx);
  int locDims[3];
  locImg->GetDimensions(locDims);

  auto rwLoc = makeOffscreenWindow(winW, winH);
  vtkNew<vtkRenderer> renLoc;
  renLoc->SetBackground(0.0, 0.0, 0.0);
  rwLoc->AddRenderer(renLoc);
  vtkSmartPointer<vtkVolume> locVol =
      geopro::render::buildVoxelI16FromImage(locImg.Get(), m.quant, cs, vmin,
                                             vmax);
  renLoc->AddVolume(locVol);
  const double locFps = benchVolumeFps(rwLoc.Get(), renLoc, frames);
  // 截图取景：局部块(256×29×162)斜俯视，呈现全分辨率细节供与概览对比。
  renLoc->ResetCamera();
  renLoc->GetActiveCamera()->Elevation(35.0);
  renLoc->GetActiveCamera()->Azimuth(25.0);
  renLoc->ResetCameraClippingRange();
  rwLoc->Render();
  const vtkIdType locNonBlack = countNonBlackPixels(rwLoc.Get(), winW, winH);
  savePng(rwLoc.Get(), (shotDir / "lod-fullres-local.png").string());
  std::cout << "[renderLOD] (b) 局部 fps=" << locFps << " 非空像素="
            << locNonBlack << " (level0 局部 " << locDims[0] << "x" << locDims[1]
            << "x" << locDims[2] << ")\n";

  // ---- (c) LOD 切换动态过渡 ----
  // 同一窗口：相机从远观（看整卷，用 level2 概览体）逐步 dolly 拉近，到一半处
  // 跨越 LOD 切换——把体从 level2 整卷换成 level0 局部体（重设 mapper 输入/相机
  // 目标），逐帧记帧耗时，标切换帧尖峰。
  std::cout << "[renderLOD] (c) LOD 切换动态过渡（" << frames << " 帧 dolly）...\n";
  auto rwTr = makeOffscreenWindow(winW, winH);
  vtkNew<vtkRenderer> renTr;
  renTr->SetBackground(0.0, 0.0, 0.0);
  rwTr->AddRenderer(renTr);

  // 远观体 = level2 概览（新建一份，避免与 (a) 共享 actor 状态）。
  vtkSmartPointer<vtkVolume> farVol =
      geopro::render::buildVoxelI16FromImage(ovImg.Get(), m.quant, cs, vmin,
                                             vmax);
  // 近观体 = level0 局部（复用 (b) 的 image）。
  vtkSmartPointer<vtkVolume> nearVol =
      geopro::render::buildVoxelI16FromImage(locImg.Get(), m.quant, cs, vmin,
                                             vmax);

  renTr->AddVolume(farVol);
  renTr->ResetCamera();  // 框住整卷（level2 与 level0 物理范围一致）
  vtkCamera* camTr = renTr->GetActiveCamera();
  camTr->Elevation(20.0);
  renTr->ResetCameraClippingRange();
  rwTr->Render();  // 预热远观

  // dolly 目标：从当前（远）拉近到局部段中心。
  double locCenter[3];
  locImg->GetCenter(locCenter);
  const int switchFrame = frames / 2;
  const double dollyPerFrame =
      std::pow(6.0, 1.0 / std::max(1, switchFrame));  // 切换前累计 dolly≈6×

  std::vector<double> frameMs(frames, 0.0);
  bool switched = false;
  double switchStallMs = 0.0;

  for (int f = 0; f < frames; ++f) {
    Stopwatch swF;
    if (f == switchFrame && !switched) {
      // —— LOD 切换那一下 ——：换体 + 把相机焦点移到局部段中心。
      renTr->RemoveVolume(farVol);
      renTr->AddVolume(nearVol);
      camTr->SetFocalPoint(locCenter[0], locCenter[1], locCenter[2]);
      renTr->ResetCameraClippingRange();
      switched = true;
    }
    // 渐进拉近（切换前 dolly 进；切换后继续推近 + 轻微环绕，逐步框满局部块）。
    camTr->Dolly(switched ? 1.04 : dollyPerFrame);
    if (switched) camTr->Azimuth(0.5);
    renTr->ResetCameraClippingRange();
    rwTr->Render();
    frameMs[f] = swF.elapsedMs();
    if (f == switchFrame) switchStallMs = frameMs[f];
    // 切换后推近一小段再截“过渡中间帧”，使局部块已明显呈现（而非切换瞬间仍很远）。
    if (f == switchFrame + (frames - switchFrame) / 3) {
      savePng(rwTr.Get(), (shotDir / "lod-transition-mid.png").string());
    }
  }

  // 过渡帧耗时统计：平均、最大、切换帧、切换帧相对邻帧的尖峰倍数。
  double sum = 0, mx = 0;
  for (double v : frameMs) {
    sum += v;
    mx = std::max(mx, v);
  }
  const double avgMs = frames > 0 ? sum / frames : 0.0;
  const double preMs =
      switchFrame > 0 ? frameMs[switchFrame - 1] : avgMs;
  const double spikeRatio = preMs > 0 ? switchStallMs / preMs : 0.0;
  // 可感知卡顿判据（绝对耗时为准，尖峰倍数仅作次级信号）：当两端帧耗时是亚毫秒
  // 时，一次性换体的 ~9ms 抖动倍数虽大但仍 <1 个 60Hz 帧(16.7ms)，人眼不可感。
  // 故：切换帧 >1 个 60Hz 帧(16.7ms)才记“轻微”，>2 帧(33ms)记“可感知卡顿”。
  constexpr double kFrame60Ms = 1000.0 / 60.0;          // 16.7ms
  const bool perceptibleStall = switchStallMs > 2.0 * kFrame60Ms;  // >33ms
  const bool minorHitch =
      !perceptibleStall && switchStallMs > kFrame60Ms;  // 16.7~33ms 轻微
  const vtkIdType trNonBlack = countNonBlackPixels(rwTr.Get(), winW, winH);

  const bool textureErr = capWin->textureError();
  vtkOutputWindow::SetInstance(nullptr);

  // 双闸：无纹理错 + 三段均渲出非空像素。
  const bool renderedNonEmpty =
      (ovNonBlack > 0) && (locNonBlack > 0) && (trNonBlack > 0);
  const bool valid = !textureErr && renderedNonEmpty;

  const double ovFpsV = valid ? ovFps : -1.0;
  const double locFpsV = valid ? locFps : -1.0;
  const bool ovInteractive = valid && ovFps >= 15.0;
  const bool locInteractive = valid && locFps >= 15.0;
  const double peak = Probe::peakMemMB();

  const char* stallTxt =
      perceptibleStall ? "可感知卡顿" : (minorHitch ? "轻微抖动(<2帧)" : "无");
  std::cout << "[renderLOD] (c) 过渡帧耗时 avg=" << avgMs << "ms max=" << mx
            << "ms 切换帧=" << switchStallMs << "ms (邻帧 " << preMs << "ms, 尖峰 "
            << spikeRatio << "×) 卡顿=" << stallTxt << "\n";

  std::cout << "\n=== renderLOD LOD-fps 探针指标 ===\n";
  std::cout << "离屏闸门          : OK\n";
  std::cout << "纹理维度错误      : " << (textureErr ? "是(!!)" : "否") << "\n";
  std::cout << "三段均渲出非空    : " << (renderedNonEmpty ? "是" : "否(!!)")
            << " (概览=" << ovNonBlack << " 局部=" << locNonBlack
            << " 过渡=" << trNonBlack << ")\n";
  std::cout << "(a) 粗层概览 fps  : "
            << (valid ? std::to_string(ovFpsV) : std::string("INVALID"))
            << " (level" << ovLevel << " " << ovNx << "x" << ovNy << "x" << ovNz
            << ") 交互级=" << (ovInteractive ? "是 ✔" : "否 ✘") << "\n";
  std::cout << "(b) 全分辨率局部fps: "
            << (valid ? std::to_string(locFpsV) : std::string("INVALID"))
            << " (level0 局部 " << locDims[0] << "x" << locDims[1] << "x"
            << locDims[2] << ") 交互级=" << (locInteractive ? "是 ✔" : "否 ✘")
            << "\n";
  std::cout << "(c) 过渡平均/最大  : " << avgMs << " / " << mx << " ms\n";
  std::cout << "    切换帧耗时    : " << switchStallMs << " ms (邻帧 " << preMs
            << " ms, 尖峰 " << spikeRatio << "×)\n";
  std::cout << "    可感知卡顿    : " << stallTxt
            << (perceptibleStall ? " ✘" : " ✔") << "  (判据:切换帧 >33ms 才记卡顿"
               "; 1 帧 60Hz=16.7ms)\n";
  std::cout << "进程峰值内存(MB)  : " << peak << "\n";
  std::cout << "截图              : " << shotDir.string()
            << " (lod-overview / lod-fullres-local / lod-transition-mid)\n";

  writeMetricLine(
      "renderLOD,dir=" + dir + ",totLevels=" + std::to_string(totLevels) +
      ",ovLevel=" + std::to_string(ovLevel) +
      ",ovDims=" + std::to_string(ovNx) + "x" + std::to_string(ovNy) + "x" +
      std::to_string(ovNz) +
      ",ovFps=" + (valid ? std::to_string(ovFpsV) : "INVALID") +
      ",ovNonBlack=" + std::to_string(ovNonBlack) +
      ",locDims=" + std::to_string(locDims[0]) + "x" +
      std::to_string(locDims[1]) + "x" + std::to_string(locDims[2]) +
      ",locFps=" + (valid ? std::to_string(locFpsV) : "INVALID") +
      ",locNonBlack=" + std::to_string(locNonBlack) +
      ",trAvgMs=" + std::to_string(avgMs) + ",trMaxMs=" + std::to_string(mx) +
      ",switchMs=" + std::to_string(switchStallMs) +
      ",switchSpike=" + std::to_string(spikeRatio) +
      ",stall=" + std::to_string(perceptibleStall ? 1 : 0) +
      ",trNonBlack=" + std::to_string(trNonBlack) +
      ",textureErr=" + std::to_string(textureErr ? 1 : 0) +
      ",valid=" + std::to_string(valid ? 1 : 0) +
      ",peakMB=" + std::to_string(peak));

  // 写 poc-results-C.md 的 LOD 段（追加，不覆盖 renderC-partitioned 段）。
  {
    const fs::path repo =
        fs::path("docs") / "superpowers" / "plans" / "poc-results-C.md";
    fs::create_directories(repo.parent_path());
    std::ofstream rf(repo.string(), std::ios::app);
    if (rf) {
      rf << "\n\n# POC-C LOD-fps 探针结果（Task 12c）\n\n";
      rf << "金字塔 store: " << dir << "（level0=" << m.nx << "x" << m.ny << "x"
         << m.nz << "，总 " << totLevels << " 层）\n\n";
      rf << "| 项 | 维度 | 结果 |\n|---|---|---|\n";
      rf << "| (a) 粗层概览 fps | level" << ovLevel << " " << ovNx << "x" << ovNy
         << "x" << ovNz << " | " << (valid ? std::to_string(ovFpsV) : "INVALID")
         << " fps " << (ovInteractive ? "(交互级)" : "(未达交互级)") << " |\n";
      rf << "| (b) 全分辨率局部 fps | level0 局部 " << locDims[0] << "x"
         << locDims[1] << "x" << locDims[2] << " | "
         << (valid ? std::to_string(locFpsV) : "INVALID") << " fps "
         << (locInteractive ? "(交互级)" : "(未达交互级)") << " |\n";
      rf << "| (c) LOD 切换过渡 | 切换帧 " << switchFrame << "/" << frames
         << " | 平均 " << avgMs << "ms，切换帧 " << switchStallMs << "ms（尖峰 "
         << spikeRatio << "×），"
         << (perceptibleStall ? "可感知卡顿"
                              : (minorHitch ? "轻微抖动" : "无可感知卡顿"))
         << " |\n\n";
      rf << "- 卡顿判据：切换帧绝对耗时 >33ms(2 个 60Hz 帧)才记可感知卡顿；"
            "16.7~33ms 记轻微抖动；亚毫秒基线下尖峰倍数大但绝对值低不算卡顿。\n";
      rf << "- 双闸：纹理维度错误=" << (textureErr ? "是" : "否")
         << "；三段均渲出非空像素=" << (renderedNonEmpty ? "是" : "否")
         << "（概览 " << ovNonBlack << " / 局部 " << locNonBlack << " / 过渡 "
         << trNonBlack << "）。\n";
      rf << "- 截图（人眼判“概览糊→拉近清晰”）：docs/superpowers/plans/poc-lod-shots/"
            "lod-overview.png、lod-fullres-local.png、lod-transition-mid.png\n";
      rf << "- 进程峰值内存: " << peak << " MB\n\n";
      rf << "## 判据结论\n";
      if (valid && ovInteractive && locInteractive && !perceptibleStall) {
        rf << "粗层概览 + 全分辨率局部【都达交互级】且切换【无不可接受卡顿】→ "
              "LOD-based C 路线钉死可行。\n";
      } else if (valid && ovInteractive && !locInteractive) {
        rf << "粗层快但全分辨率局部仍慢 → VTK 体绘制有真实天花板，记录，"
              "评估 OpenVDS/自建 GL。\n";
      } else if (valid && perceptibleStall) {
        rf << "两端 fps 可接受但切换卡顿明显（切换帧 " << switchStallMs
           << "ms）→ 为后续 morphing/淡入提供依据。\n";
      } else if (!valid) {
        rf << "双闸未过（纹理错或空渲染）→ 数字不可信，如实标 INVALID。\n";
      } else {
        rf << "部分达标，详见上表。\n";
      }
      rf << "\n**最低配未验声明**：本探针仅在本机（RTX 3060）跑得上限数字，"
            "最低配机器未验证，需用户在目标机跑或提供型号。\n";
    }
    std::cout << "[renderLOD] 报告追加写入 " << repo.string() << "\n";
  }

  return valid ? 0 : 1;
}

void usage() {
  std::cerr << "gpr_poc —— POC-B headless 度量 CLI\n"
               "  gpr_poc build <dir> [--line 001] [--cellXY 0.2] "
               "[--cellZ 0.05] [--out <storeDir>] [--levels 2]\n"
               "  gpr_poc load  <storeDir>\n"
               "  gpr_poc selftest\n"
               "  gpr_poc offscreen-smoke\n"
               "  gpr_poc renderB <storeDir> [--frames 120]\n"
               "  gpr_poc renderC <storeDir> [--budget 64] [--frames 120]\n"
               "  gpr_poc renderC-partitioned <storeDir> [--frames 120]\n"
               "  gpr_poc renderLOD <storeDir> [--frames 120]\n";
}

}  // namespace

int main(int argc, char** argv) {
  if (argc < 2) {
    usage();
    return 2;
  }
  const std::string cmd = argv[1];
  try {
    if (cmd == "build") return cmdBuild(argc, argv);
    if (cmd == "load") return cmdLoad(argc, argv);
    if (cmd == "selftest") return cmdSelftest();
    if (cmd == "offscreen-smoke") return cmdOffscreenSmoke();
    if (cmd == "renderB") return cmdRenderB(argc, argv);
    if (cmd == "renderC") return cmdRenderC(argc, argv);
    if (cmd == "renderC-partitioned")
      return cmdRenderCPartitioned(argc, argv);
    if (cmd == "renderLOD") return cmdRenderLOD(argc, argv);
  } catch (const std::exception& e) {
    std::cerr << "错误: " << e.what() << "\n";
    return 1;
  }
  usage();
  return 2;
}