sar3d: fixed default scene window X [-2,2], Y [-0.5,2], Z [0,8] m

The grid bounds are now explicit config (sar3d_x/y/z_min/max_m) instead of
being derived from the antenna span and the depth gate. The depth mask
still zeroes voxels outside [min_depth, max_depth], so the wider Z window
just shows the masked wedge. The bound fields are double: the parity tool
recomputes the notebook-derived bounds (notebook_grid_bounds) and passes
them through the live config, and a float roundtrip shifted the axes by
~1e-7 m, flipping depth-mask knife-edge voxels against the reference npz.
Parity on frames 0034-0036 unchanged (1.3e-6, axes bit-match).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Makefile

@@ -145,11 +145,14 @@ all: $(TARGETS)
 
 kamil_adc_collector: $(BIN_DIR)/kamil_adc_collector
 
-# On-demand parity check of the sar3d processor against the gpr_cpp
-# standalone_bp example (same .npy dataset). Not part of `all`.
+# On-demand parity check of the sar3d processor against the reference notebook
+# volumes (scripts/sar3d_volume_compare.py diffs the dumped volume against a
+# *_volume.npz export). Not part of `all`.
 SAR3D_PARITY_OBJS := $(addprefix $(BUILD_DIR)/, \
 	data_acq_and_processing/processing/tools/sar3d_parity_main.o \
-	data_acq_and_processing/processing/processors/src/sar3d_processor.o)
+	data_acq_and_processing/processing/processors/src/sar3d_processor.o \
+	data_acq_and_processing/common_cpp/config/src/run_config.o \
+	data_acq_and_processing/common_cpp/ipc/src/shared_types.o)
 
 sar3d_parity: $(BIN_DIR)/sar3d_parity
 

data_acq_and_processing/gpr_cpp/README.md

@@ -1,23 +1,24 @@
 # gpr_cpp — C++/CUDA счёт-ядро MIMO GPR 3D SAR
 
-> Вендорено из репозитория VNA_AlGo (`gpr_cpp/`, коммит `544dc59`); python-биндинги
-> (`bindings/`, pybind11) при переносе не взяты — сборка здесь только C++
-> (`GPR_BUILD_PYTHON=OFF`). Собирается из корневого Makefile radar_system
+> Инфраструктура (сборка, NVRTC, PocketFFT, npy) вендорена из репозитория
+> VNA_AlGo (`gpr_cpp/`, коммит `544dc59`) без python-биндингов
+> (`GPR_BUILD_PYTHON=OFF`). АЛГОРИТМ с сентября 2026 переписан: это порт
+> эталонного ноутбука `Multi_antenna_3D_clean_v1.ipynb` (ячейки 0 и 4), а не
+> старого `gpr_core`/Horns_3D. Паритет с ноутбуком проверяется инструментом
+> `sar3d_parity` + `python_app/scripts/sar3d_volume_compare.py` по
+> референсным `*_volume.npz`. Собирается из корневого Makefile radar_system
 > (цель `$(GPR_CPP_LIB)` вызывает `build_cmake.sh`).
 
-Порт вычислительного ядра из Python/CuPy (`gpr_core`) в самостоятельную
-C++/CUDA библиотеку для интеграции в C++ проект.
-Алгоритм — см. `ALGORITHM.md` в репозитории VNA_AlGo.
+## Что внутри
 
-## Что внутри (граница C++ ↔ Python)
-
-**В C++ (`libgpr_core_cpp`):** геометрия, препроцессинг A-сканов (bg-subtract,
-фазовая коррекция движения, k0-вставка + Ханн + IFFT через PocketFFT, нормировка),
-3D SAR апертурный объём (одно CUDA-ядро на всю апертуру). Возвращает **сырые**
-`acc`(complex64)/`cnt`(float32).
-
-**Остаётся в Python:** загрузка `.npy`, финализация объёма (abs+gate+gaussian+
-нормировка), CLEAN-детекция, трек, оценка скорости, plotly.
+**В C++ (`libgpr_core_cpp`):** сегментированный тайминг кадра по per-trace
+меткам захвата (`compute_pair_timestamps`), препроцессинг A-сканов (bg-subtract,
+полная фазовая коррекция движения intra+pair до IFFT, k0-вставка + Ханн + IFFT
+через PocketFFT, нормировка пар), полнообластная ground-сетка
+(`build_full_region_grid`), 3D SAR апертурный объём (одно CUDA-ядро на всю
+апертуру; геометрия в double, A-сканы float32) и финализация
+(`finish_volume`: |acc/cnt| → маска глубины → нормировка без Y-краёв → гаусс
+по (Z,X) → нормировка; возвращает объём [0,1] + level).
 
 ## Ключевое решение: CUDA без nvcc (NVRTC + Driver API)
 
@@ -42,14 +43,14 @@ libnvrtc прописан — бинарники/модуль грузятся
 ## Публичный C++ API ([include/gpr/](include/gpr/))
 
 ```cpp
-#include "gpr/config.hpp"    // GeometryConfig/BPConfig/MotionConfig/SARConfig (дефолты = config.py)
-#include "gpr/ascan.hpp"     // preprocess_snapshot(...) -> AScanSet
-#include "gpr/sar3d.hpp"     // SarEngine::aperture_volume(frames, grid) -> VolumeRaw{acc,cnt}
-#include "gpr/geometry.hpp"  // frame_offsets, frame_weights, build_grid_scene, point_to_scene
+#include "gpr/config.hpp"    // GeometryConfig/BPConfig/MotionConfig/SARConfig (дефолты = ноутбук)
+#include "gpr/ascan.hpp"     // compute_pair_timestamps, preprocess_snapshot(...) -> AScanSet
+#include "gpr/sar3d.hpp"     // SarEngine::aperture_volume -> VolumeRaw; finish_volume -> FinishedVolume
+#include "gpr/geometry.hpp"  // frame_offsets, frame_weights, build_full_region_grid
 
 gpr::SarEngine engine(geom, bp, mot, sar);          // компилит ядро под устройство
-auto vol = engine.aperture_volume(frames, grid);    // один запуск ядра на апертуру
-// vol.acc / vol.cnt — сырые; финализацию (gaussian/normalize) делает вызывающий
+auto raw = engine.aperture_volume(frames, grid);    // один запуск ядра на апертуру
+auto fin = gpr::finish_volume(raw, grid, mot.pitch_deg, bp, sar);  // объём [0,1] + level
 ```
 Пример без Python: [examples/standalone_bp.cpp](examples/standalone_bp.cpp) — грузит
 S21 из `.npy`, строит объём, печатает максимум.
@@ -67,16 +68,13 @@ a = _gpr_cpp.preprocess_snapshot(s21, freq, background, mot, geom, bp, True)  #
 acc, cnt = eng.aperture_volume([(a, dx, dy, dz, w), ...], gx, gy, gz)         # numpy
 ```
 
-## Паритет (проверено на x86 T600)
+## Паритет с ноутбуком (проверено на x86 T600, датасет 20260922_mine_fast_true)
 
-| Этап | Порог | Факт |
-|---|---|---|
-| config дефолты vs dataclass | точно | OK |
-| A-сканы raw H vs numpy (float64) | <1e-9 | 1.7e-12 |
-| A-сканы норм. H (FFT-ULP через медиану) | <1e-4 | 2.5e-6 |
-| **3D объём C++ ядро vs cupy** (float32) | <5e-3 | **3e-7, argmax идентичен** |
-
-Тесты: [../tests/test_parity_cpp.py](../tests/test_parity_cpp.py) (C++ vs cupy).
+Кадры 0034/0035/0036 против референсных `*_volume.npz` ноутбука
+(`make sar3d_parity` + `python_app/scripts/sar3d_volume_compare.py`):
+max|Δ| ≤ 1.6e-6 вне ножевой границы маски глубины, argmax и level совпадают
+(level rel ≤ 7e-7). Python-эталон той же цепочки:
+`python_app/scripts/sar3d_reference_check.py`.
 
 ## Статус
 

data_acq_and_processing/gpr_cpp/examples/standalone_bp.cpp

@@ -95,7 +95,20 @@ int main(int argc, char** argv) {
         }
     };
 
-    auto ts = compute_pair_timestamps(mot, geom);
+    // Демо без чтения meta.json: равномерные сегменты по интервалу между
+    // соседними снапшотами (реальные per-trace окна читает sar3d_parity).
+    double acq_s = 0.0;
+    {
+        std::vector<double> stamps;
+        for (auto& s : snaps) {
+            auto pos = s.rfind("_ns");
+            if (pos != std::string::npos) stamps.push_back(std::stod(s.substr(pos + 3)) * 1e-9);
+        }
+        std::sort(stamps.begin(), stamps.end());
+        if (stamps.size() >= 2) acq_s = stamps[1] - stamps[0];
+    }
+    auto windows = uniform_tx_windows(geom.output_idx, acq_s, mot.tx_per_segment);
+    auto ts = compute_pair_timestamps(windows, acq_s, geom.output_idx, geom, mot);
     SarEngine engine(geom, bp, mot, sar);
 
     // держим AScanSet живыми
@@ -122,22 +135,23 @@ int main(int argc, char** argv) {
         frames.push_back(ApertureFrame{a.get(), offs[f], wts[f]});
     }
 
-    // центр сетки: под центральным снапшотом на глубине 5.6 м (сцена)
-    auto center_scene = point_to_scene(0.0, 0.0, 5.6, mot, sar);
-    Grid grid = build_grid_scene(center_scene, sar);
+    // ноутбучные границы сцены (размах антенн ±2 м, полоса глубин) вместо
+    // фиксированного окна по умолчанию
+    notebook_grid_bounds(geom, bp, sar, mot.pitch_deg, 2.0,
+                         sar.x_min_m, sar.x_max_m, sar.z_min_m, sar.z_max_m);
+    Grid grid = build_full_region_grid(sar);
     printf("сетка %dx%dx%d, счёт объёма...\n", grid.NX(), grid.NY(), grid.NZ());
 
     VolumeRaw vol = engine.aperture_volume(frames, grid);
+    FinishedVolume fin = finish_volume(vol, grid, mot.pitch_deg, bp, sar);
 
-    // финализация: |acc/cnt|, максимум
     double vmax = 0; size_t amax = 0;
-    for (size_t i = 0; i < vol.acc.size(); ++i) {
-        double v = std::abs(vol.acc[i]) / (vol.cnt[i] + 1e-12);
-        if (v > vmax) { vmax = v; amax = i; }
-    }
+    for (size_t i = 0; i < fin.volume.size(); ++i)
+        if (fin.volume[i] > vmax) { vmax = fin.volume[i]; amax = i; }
     int NX = vol.NX, NZ = vol.NZ;
     int ix = amax % NX, iz = (amax / NX) % NZ, iy = amax / ((size_t)NX * NZ);
-    printf("МАКСИМУМ объёма: |acc/cnt|=%.4e в вокселе (ix=%d,iy=%d,iz=%d)\n", vmax, ix, iy, iz);
+    printf("МАКСИМУМ объёма: %.4f (level %.4e) в вокселе (ix=%d,iy=%d,iz=%d)\n",
+           vmax, fin.level, ix, iy, iz);
     printf("  X=%.1f см, Y=%.1f см, Z=%.1f см\n",
            grid.x[ix] * 100, grid.y[iy] * 100, grid.z[iz] * 100);
     printf("OK: автономный C++ путь отработал без Python.\n");

data_acq_and_processing/gpr_cpp/include/gpr/ascan.hpp

@@ -1,4 +1,6 @@
-// Препроцессинг S21 → комплексные A-сканы (host, float64). Порт preprocess.py.
+// Препроцессинг S21 → комплексные A-сканы (host, float64).
+// Порт ячейки 0 Multi_antenna_3D_clean_v1.ipynb (segmented_interleaved тайминг,
+// frequency-domain коррекция движения, k0-IFFT, нормировка пар).
 #pragma once
 #include <map>
 #include <string>
@@ -12,15 +14,34 @@ namespace gpr {
 
 using PairKey = std::pair<int, int>;
 
-// Тайминги пары (нужны slot/n_slots/t_center для фазовой коррекции).
+// Тайминг пары в модели segmented_interleaved: пара живёт в сегменте своего Tx
+// (окно [segment_start, segment_start+segment_time] от начала акквизиции),
+// внутри сегмента передатчики чередуются по частоте (slot из n_slots).
 struct PairTiming {
     int tx_interleave_slot = 0;
     int tx_interleave_slots = 1;
-    double t_center_s = 0.0;
+    double segment_start_s = 0.0;
+    double segment_time_s = 0.0;
+    double t_center_s = 0.0;   // центр окна сегмента
+    // Грубый межсегментный сдвиг: dtau центра сегмента относительно центра
+    // акквизиции кадра (2*dz/v, dz = dir*speed*dt*cos(look)).
+    double dtau_pair_s = 0.0;
 };
 using PairTimestamps = std::map<PairKey, PairTiming>;
 
-PairTimestamps compute_pair_timestamps(const MotionConfig&, const GeometryConfig&);
+// Окна захвата по Tx (output_pos → [start,end] сек от начала акквизиции).
+using TxWindows = std::map<int, std::pair<double, double>>;
+
+// Fallback без per-trace меток: сегменты по tx_per_segment штук делят кадр
+// [0, acq_s] поровну (медианная модель ноутбука compute_pair_timestamps).
+TxWindows uniform_tx_windows(const std::vector<int>& tx_order, double acq_s, int tx_per_segment);
+
+// Порт compute_pair_timestamps_for_snapshot: Tx с одинаковым окном — один
+// сегмент (сегменты по возрастанию окна), слот — порядок в tx_order (полный
+// аппаратный порядок из run.combos; выбранные каналы его не меняют).
+PairTimestamps compute_pair_timestamps(const TxWindows& tx_windows_s, double acq_s,
+                                       const std::vector<int>& tx_order,
+                                       const GeometryConfig&, const MotionConfig&);
 
 // A-сканы одного снапшота, общая ось времени у всех пар (assert).
 struct AScanSet {
@@ -32,7 +53,7 @@ struct AScanSet {
     const c128* pair_ptr(int p) const { return H.data() + (size_t)p * n_fft; }
 };
 
-// Порт compute_pair_ascans: bg-subtract → phase correction → k0-IFFT → норма.
+// Порт sar_pair_ascans: bg-subtract → phase correction → k0-IFFT → норма.
 // s21/freq/background — по ключам (i_tx,i_rx); background может быть nullptr.
 AScanSet preprocess_snapshot(
     const std::map<PairKey, std::vector<c128>>& s21,
@@ -42,10 +63,12 @@ AScanSet preprocess_snapshot(
     const PairTimestamps&, bool normalize = true);
 
 // Отдельные шаги (для тестов/переиспользования):
+// Полная фазовая модель движения до IFFT: dtau_intra(f) + dtau_pair одной
+// фазой exp(sign*j*2*pi*f*dtau_total) (порт apply_intra_sweep_phase_correction).
 std::vector<c128> apply_intra_sweep_phase_correction(
     const std::vector<c128>& s21, const std::vector<double>& freq_full,
     const PairTiming&, const MotionConfig&, double wave_speed,
-    double f_start, double f_stop, int n_slots_default);
+    double f_start, double f_stop);
 
 // Возвращает t (ось), h (комплексный A-скан). n_fft через out-параметр.
 void compute_ascan_bp(const std::vector<c128>& s21, const std::vector<double>& freq,

data_acq_and_processing/gpr_cpp/include/gpr/config.hpp

@@ -1,4 +1,5 @@
-// Конфигурация счёт-ядра. Дефолты 1:1 повторяют gpr_core/config.py.
+// Конфигурация счёт-ядра. Дефолты 1:1 повторяют Multi_antenna_3D_clean_v1.ipynb
+// (ячейки 0 и 4) — эталонный ноутбук 3D SAR.
 #pragma once
 #include <algorithm>
 #include <array>
@@ -11,13 +12,13 @@ namespace gpr {
 
 struct GeometryConfig {
     std::vector<int> input_idx{0, 1, 2, 3};
-    std::vector<int> output_idx{0, 1};
-    // (x,y,z) метры; порядок по физическим индексам как в config.py
-    std::vector<std::array<double, 3>> tx_positions{{0.785, 0.0, 0.0}, {-0.785, 0.0, 0.0}};
+    std::vector<int> output_idx{0, 1, 2, 3};
+    // (x,y,z) метры; порядок соответствует input_idx/output_idx
+    std::vector<std::array<double, 3>> tx_positions{
+        {-0.88, 0.0, 0.0}, {0.89, 0.0, 0.0}, {-0.33, 0.445, 0.0}, {0.33, 0.445, 0.0}};
     std::vector<std::array<double, 3>> rx_positions{
-        {-0.42, 0.0, 0.0}, {0.48, 0.0, 0.0}, {0.18, 0.0, 0.0}, {-0.12, 0.0, 0.0}};
+        {0.23, 0.0, 0.0}, {0.54, 0.15, 0.0}, {-0.55, 0.3, 0.0}, {-0.23, 0.0, 0.0}};
     double eps_r = 1.0;
-    double bp_plane_y = 0.0;
 
     // --- производные (заполняются finalize) ---
     std::vector<double> x_tx, y_tx, z_tx;   // отсортированы по phys-индексу
@@ -50,53 +51,64 @@ struct GeometryConfig {
 };
 
 struct BPConfig {
-    double f_start = 20 * 1e8;   // 2 ГГц
+    double f_start = 2.0e9;
     double f_stop = 6.0e9;
-    double min_depth = 3.2;
-    double max_depth = 10.0;
-    double grid_x_pad = 2.0;
-    double z_min = 0.1;
-    int nx = 300;
-    int nz = 300;
+    double min_depth = 2.5;
+    double max_depth = 8.0;
     int oversample = 8;
-    bool window = true;
+    bool window = true;                 // окно Ханна на полосе
     bool pair_normalize = true;
     double pair_norm_percentile = 50.0;
-    double pair_norm_depth_min = 3.2;
-    double pair_norm_depth_max = 10.0;
+    double pair_norm_depth_min = 2.5;
+    double pair_norm_depth_max = 8.0;
     double pair_norm_eps = 1e-15;
-    double comp_range_power = 0.25;
+    // Компенсация геометрического расхождения 1/(Rtx*Rrx), нормированная на
+    // точку под виртуальным центром пары на глубине comp_ref_depth НА ПЛОСКОСТИ
+    // ЗЕМЛИ (высота (-ground_depth + d*sin p)/cos p).
+    double comp_range_power = 0.3;
     double comp_weight_max = 5.0;
     double comp_ref_depth = 5.0;
-    double smooth_sigma = 1.5;
-    bool bg_subtract = true;
-    double cf_eps = 1e-12;
 };
 
 struct MotionConfig {
-    double speed_m_s = 0.854;
-    double look_angle_deg = 6.0;
-    double pitch_deg = 6.0;
-    double tx_sweep_time_s = 0.079;
-    double tx_switch_time_s = 1e-5;
-    // frequency_timing_mode всегда interleaved_tx_by_frequency
+    double speed_m_s = 1.45;
+    double look_angle_deg = 17.5;
+    // Наклон рамы для ground-координат и смещений кадров; в ноутбуке
+    // BP_GROUND_PITCH_DEG=None означает «равен look_angle».
+    double pitch_deg = 17.5;
     double direction_sign = +1.0;
+    // Внутрисвиповая (по частоте) фазовая коррекция движения.
     bool apply_freq_phase_correction = true;
+    // Грубая межсегментная коррекция (dtau сегмента относительно центра кадра),
+    // вносится той же фазовой моделью до IFFT (режим 'frequency_domain').
+    bool apply_pair_motion_correction = true;
     double motion_phase_sign = -1.0;
-
-    double frame_time_s(int n_tx) const {
-        return n_tx * tx_sweep_time_s + std::max(0, n_tx - 1) * tx_switch_time_s;
-    }
+    // Fallback-разбиение кадра на сегменты, когда per-trace меток захвата нет.
+    int tx_per_segment = 2;
 };
 
 struct SARConfig {
-    int frame_radius = 6;
-    double edge_frame_weight = 0.75;
-    std::array<double, 3> half_width_m{1.8, 1.0, 1.8};
-    std::array<int, 3> grid{150, 120, 160};   // NX, NY, NZ
-    bool coord_mode_ground = true;             // 'ground' vs 'radar_local'
+    int frame_radius = 0;                 // 0 = один снимок (физическая вертикальная база)
+    double edge_frame_weight = 0.8;       // вес крайнего кадра апертуры
+    bool coord_mode_ground = true;        // 'ground' vs 'radar_local'
     bool use_compensation = true;
-    bool normalize_each_frame = true;
+    // Высота НИЖНЕГО ряда антенн над землёй; Y=0 сцены — поверхность земли.
+    double ground_depth_m = 1.25;
+    // Явные границы сцены [м]: X — поперёк, Y — высота над землёй, Z — вдаль.
+    // Воксели с радарной глубиной вне [min_depth, max_depth] всё равно
+    // обнуляются маской, так что Z-окно шире полосы глубин безопасно.
+    double x_min_m = -2.0;
+    double x_max_m = 2.0;
+    double y_min_m = -0.5;
+    double y_max_m = 2.0;
+    double z_min_m = 0.0;
+    double z_max_m = 8.0;
+    std::array<double, 3> voxel_m{0.025, 0.02, 0.02};   // шаг (X, Y, Z)
+    // Финализация объёма: нормировка без верхних/нижних norm_edge_margin_m по Y
+    // (там ложный гребень постоянной дальности), сглаживание по (Z,X) в каждом
+    // Y-плане.
+    double norm_edge_margin_m = 0.3;
+    double smooth_sigma = 1.5;
 };
 
 }  // namespace gpr

data_acq_and_processing/gpr_cpp/include/gpr/geometry.hpp

@@ -18,10 +18,23 @@ std::vector<FrameOffset> frame_offsets(const std::vector<double>& frame_times, d
 // Косинусные веса апертуры (порт sar_frame_weights); dt берётся относительно центра.
 std::vector<double> frame_weights(const std::vector<FrameOffset>& offsets, double edge_weight);
 
-// Сетка вокруг точки сцены (порт sar_build_grid_scene).
-Grid build_grid_scene(const std::array<double, 3>& center_scene, const SARConfig&);
+// Высота плоскости земли (радарные оси) на радарной глубине depth:
+// y = (-ground_depth + depth*sin(pitch)) / cos(pitch)  (порт bp_plane_y).
+double ground_plane_y(double depth, double ground_depth_m, double pitch_rad);
 
-// point_to_scene: оси радара → сцена (поворот на pitch, если ground).
+// Сетка сцены по явным границам SARConfig (x/y/z_min..max, шаг voxel_m);
+// оси — linspace c числом точек round(span/step)+1, как в ноутбуке.
+Grid build_full_region_grid(const SARConfig&);
+
+// Ноутбучные границы (sar_build_full_region_grid): X — размах антенн ±margin,
+// Z — полоса, покрывающая радарные глубины [min_depth, max_depth] на всех
+// высотах Y. Используются parity-инструментом для сверки с эталонными npz.
+void notebook_grid_bounds(const GeometryConfig&, const BPConfig&, const SARConfig&,
+                          double pitch_deg, double x_margin_m,
+                          double& x_min, double& x_max, double& z_min, double& z_max);
+
+// point_to_scene: оси радара → сцена (поворот на pitch + подъём на ground_depth,
+// если ground; Y=0 сцены — земля).
 std::array<double, 3> point_to_scene(double x, double y, double z, const MotionConfig&,
                                      const SARConfig&);
 

data_acq_and_processing/gpr_cpp/include/gpr/sar3d.hpp

@@ -19,13 +19,24 @@ struct ApertureFrame {
     double weight;
 };
 
-// Сырой результат объёма (Python делает finish: abs+gate+gaussian+normalize).
+// Сырой результат объёма (до finish_volume).
 struct VolumeRaw {
     std::vector<c64> acc;    // [NY*NZ*NX]
     std::vector<float> cnt;  // [NY*NZ*NX]
     int NX = 0, NY = 0, NZ = 0;
 };
 
+// Финальный объём (порт sar_finish_volume): |acc/cnt| → маска глубины →
+// нормировка по Y-планам без краёв → гаусс по (Z,X) → нормировка → clip.
+struct FinishedVolume {
+    std::vector<float> volume;   // [NY*NZ*NX], нормирован на ~[0,1]
+    double level = 0.0;          // масштаб, на который поделён объём
+    int NX = 0, NY = 0, NZ = 0;
+};
+
+FinishedVolume finish_volume(const VolumeRaw&, const Grid&, double pitch_deg,
+                             const BPConfig&, const SARConfig&);
+
 class SarEngine {
    public:
     SarEngine(const GeometryConfig&, const BPConfig&, const MotionConfig&, const SARConfig&);

data_acq_and_processing/gpr_cpp/src/ascan.cpp

@@ -9,19 +9,69 @@
 
 namespace gpr {
 
-PairTimestamps compute_pair_timestamps(const MotionConfig& mot, const GeometryConfig& geom) {
-    // interleaved_tx_by_frequency: слот Tx = порядок в отсортированном output_idx.
-    std::vector<int> tx = geom.output_idx, rx = geom.input_idx;
-    std::sort(tx.begin(), tx.end());
-    std::sort(rx.begin(), rx.end());
-    int n_slots = (int)tx.size();
-    double frame_center = 0.5 * n_slots * mot.tx_sweep_time_s;
+TxWindows uniform_tx_windows(const std::vector<int>& tx_order, double acq_s, int tx_per_segment) {
+    TxWindows windows;
+    if (tx_order.empty()) return windows;
+    int per_seg = std::max(1, tx_per_segment);
+    int n_seg = std::max(1, (int)((tx_order.size() + per_seg - 1) / per_seg));
+    double t_seg = acq_s / n_seg;
+    for (size_t k = 0; k < tx_order.size(); ++k) {
+        int seg = (int)k / per_seg;
+        windows[tx_order[k]] = {seg * t_seg, seg * t_seg + t_seg};
+    }
+    return windows;
+}
+
+PairTimestamps compute_pair_timestamps(const TxWindows& tx_windows_s, double acq_s,
+                                       const std::vector<int>& tx_order,
+                                       const GeometryConfig& geom, const MotionConfig& mot) {
+    // Аппаратный порядок: сперва tx_order (run.combos), затем прочие ключи окон.
+    std::vector<int> hw_order;
+    for (int tx : tx_order)
+        if (tx_windows_s.count(tx) && std::find(hw_order.begin(), hw_order.end(), tx) == hw_order.end())
+            hw_order.push_back(tx);
+    for (auto& [tx, _] : tx_windows_s)
+        if (std::find(hw_order.begin(), hw_order.end(), tx) == hw_order.end())
+            hw_order.push_back(tx);
+
+    // Tx с одинаковым окном — один сегмент; сегменты по возрастанию окна,
+    // слот — порядок появления в hw_order (порт группировки ноутбука).
+    std::vector<std::pair<std::pair<double, double>, std::vector<int>>> groups;
+    for (int tx : hw_order) {
+        auto win = tx_windows_s.at(tx);
+        auto found = std::find_if(groups.begin(), groups.end(),
+                                  [&](const auto& g) { return g.first == win; });
+        if (found == groups.end())
+            groups.push_back({win, {tx}});
+        else
+            found->second.push_back(tx);
+    }
+    std::sort(groups.begin(), groups.end(),
+              [](const auto& a, const auto& b) { return a.first < b.first; });
+
+    struct SegSlot { double t0, t1; int slot, n_slots; };
+    std::map<int, SegSlot> seg_of;
+    for (auto& [win, txs] : groups)
+        for (size_t slot = 0; slot < txs.size(); ++slot)
+            seg_of[txs[slot]] = {win.first, win.second, (int)slot, (int)txs.size()};
+
+    double cos_theta = std::cos(mot.look_angle_deg * M_PI / 180.0);
+    double t_ref = 0.5 * acq_s;
     PairTimestamps ts;
-    for (int slot = 0; slot < (int)tx.size(); ++slot) {
-        int i_tx = slot;  // индекс в отсортированном == логический
-        for (int j = 0; j < (int)rx.size(); ++j) {
-            ts[{i_tx, j}] = PairTiming{slot, n_slots, frame_center};
-        }
+    for (int i = 0; i < geom.n_tx; ++i) {
+        auto it = seg_of.find(geom.output_idx[i]);
+        if (it == seg_of.end())
+            throw std::runtime_error("pair timestamps: нет окна для Tx");
+        const SegSlot& s = it->second;
+        PairTiming timing;
+        timing.tx_interleave_slot = s.slot;
+        timing.tx_interleave_slots = s.n_slots;
+        timing.segment_start_s = s.t0;
+        timing.segment_time_s = s.t1 - s.t0;
+        timing.t_center_s = 0.5 * (s.t0 + s.t1);
+        double dz = mot.direction_sign * mot.speed_m_s * (timing.t_center_s - t_ref) * cos_theta;
+        timing.dtau_pair_s = 2.0 * dz / geom.v;
+        for (int j = 0; j < geom.n_rx; ++j) ts[{i, j}] = timing;
     }
     return ts;
 }
@@ -37,7 +87,7 @@ static double median_of_diff(const std::vector<double>& f) {
 std::vector<c128> apply_intra_sweep_phase_correction(
     const std::vector<c128>& s21, const std::vector<double>& freq_full,
     const PairTiming& info, const MotionConfig& cfg, double wave_speed,
-    double f_start, double f_stop, int n_slots_default) {
+    double f_start, double f_stop) {
     std::vector<c128> out = s21;
     int n_full = (int)freq_full.size();
     // индексы в полосе
@@ -47,21 +97,23 @@ std::vector<c128> apply_intra_sweep_phase_correction(
     if (idx.empty()) throw std::runtime_error("phase corr: пустая полоса");
     if (n_full < 2) return out;
 
-    int n_slots = info.tx_interleave_slots ? info.tx_interleave_slots : n_slots_default;
+    // Сегмент проходит все n_full частот за своё окно; на каждой частоте по
+    // очереди работают его n_slots передатчиков, поэтому событие частоты i для
+    // слота slot имеет номер (n_slots*i + slot) из (n_slots*n_full).
+    int n_slots = std::max(1, info.tx_interleave_slots);
     int slot = info.tx_interleave_slot;
-    double dt_base = cfg.tx_sweep_time_s / (n_full - 1);
-    // t_all по всему свипу; t_center — середина ПОЛНОГО свипа (как в preprocess.py)
-    double t_first = (double)(n_slots * 0 + slot) * dt_base;
-    double t_last = (double)(n_slots * (n_full - 1) + slot) * dt_base;
-    double t_center = 0.5 * (t_first + t_last);
-    double theta = cfg.look_angle_deg * M_PI / 180.0;
-    if (!cfg.apply_freq_phase_correction) return out;
+    double dt_event = info.segment_time_s / ((double)n_slots * n_full);
+    double cos_theta = std::cos(cfg.look_angle_deg * M_PI / 180.0);
+    double pair_dtau = cfg.apply_pair_motion_correction ? info.dtau_pair_s : 0.0;
+    if (!cfg.apply_freq_phase_correction && pair_dtau == 0.0) return out;
     for (int i : idx) {
-        double t_abs = (double)(n_slots * i + slot) * dt_base;
-        double dt_intra = t_abs - t_center;
-        double delta_range = cfg.direction_sign * cfg.speed_m_s * std::cos(theta) * dt_intra;
-        double dtau = 2.0 * delta_range / wave_speed;
-        double phi = 2.0 * M_PI * freq_full[i] * dtau;
+        double dtau_intra = 0.0;
+        if (cfg.apply_freq_phase_correction) {
+            double t_abs = info.segment_start_s + (double)(n_slots * i + slot) * dt_event;
+            double dt_intra = t_abs - info.t_center_s;
+            dtau_intra = 2.0 * cfg.direction_sign * cfg.speed_m_s * cos_theta * dt_intra / wave_speed;
+        }
+        double phi = 2.0 * M_PI * freq_full[i] * (dtau_intra + pair_dtau);
         double ang = cfg.motion_phase_sign * phi;
         out[i] *= c128(std::cos(ang), std::sin(ang));
     }
@@ -163,7 +215,7 @@ AScanSet preprocess_snapshot(
         auto itt = ts.find(key);
         if (itt != ts.end())
             s = apply_intra_sweep_phase_correction(s, fr, itt->second, mot, geom.v,
-                                                   bp.f_start, bp.f_stop, geom.n_tx);
+                                                   bp.f_start, bp.f_stop);
         std::vector<double> t;
         std::vector<c128> h;
         compute_ascan_bp(s, fr, bp.f_start, bp.f_stop, bp.window, bp.oversample, t, h);

data_acq_and_processing/gpr_cpp/src/geometry.cpp

@@ -1,5 +1,6 @@
 #include "gpr/geometry.hpp"
 
+#include <algorithm>
 #include <cmath>
 
 namespace gpr {
@@ -48,21 +49,56 @@ std::array<double, 3> point_to_scene(double x, double y, double z, const MotionC
                                      const SARConfig& sar) {
     if (!sar.coord_mode_ground) return {x, y, z};
     double p = mot.pitch_deg * M_PI / 180.0;
-    double ys = y * std::cos(p) - z * std::sin(p);
+    double ys = y * std::cos(p) - z * std::sin(p) + sar.ground_depth_m;
     double zs = y * std::sin(p) + z * std::cos(p);
     return {x, ys, zs};
 }
 
-Grid build_grid_scene(const std::array<double, 3>& c, const SARConfig& sar) {
-    Grid g;
-    auto lin = [](double a, double b, int n, std::vector<double>& out) {
+double ground_plane_y(double depth, double ground_depth_m, double pitch_rad) {
+    return (-ground_depth_m + depth * std::sin(pitch_rad)) / std::cos(pitch_rad);
+}
+
+Grid build_full_region_grid(const SARConfig& sar) {
+    // numpy: linspace(lo, hi, max(round(span/step)+1, 2)); nearbyint = банковское
+    // округление python round()
+    auto axis = [](double lo, double hi, double step, std::vector<double>& out) {
+        int n = std::max((int)std::nearbyint((hi - lo) / step) + 1, 2);
         out.resize(n);
-        for (int i = 0; i < n; ++i) out[i] = a + (b - a) * i / (double)(n - 1);
+        for (int i = 0; i < n; ++i) out[i] = lo + (hi - lo) * i / (double)(n - 1);
     };
-    lin(c[0] - sar.half_width_m[0], c[0] + sar.half_width_m[0], sar.grid[0], g.x);
-    lin(c[1] - sar.half_width_m[1], c[1] + sar.half_width_m[1], sar.grid[1], g.y);
-    lin(c[2] - sar.half_width_m[2], c[2] + sar.half_width_m[2], sar.grid[2], g.z);
+    Grid g;
+    axis(sar.x_min_m, sar.x_max_m, sar.voxel_m[0], g.x);
+    axis(sar.y_min_m, sar.y_max_m, sar.voxel_m[1], g.y);
+    axis(sar.z_min_m, sar.z_max_m, sar.voxel_m[2], g.z);
     return g;
 }
 
+void notebook_grid_bounds(const GeometryConfig& geom, const BPConfig& bp, const SARConfig& sar,
+                          double pitch_deg, double x_margin_m,
+                          double& x_min, double& x_max, double& z_min, double& z_max) {
+    double x_lo = geom.x_tx[0], x_hi = geom.x_tx[0];
+    for (double x : geom.x_tx) { x_lo = std::min(x_lo, x); x_hi = std::max(x_hi, x); }
+    for (double x : geom.x_rx) { x_lo = std::min(x_lo, x); x_hi = std::max(x_hi, x); }
+    x_min = x_lo - x_margin_m;
+    x_max = x_hi + x_margin_m;
+
+    // Радарная z линейна по Z сцены при фиксированной высоте:
+    // z_rad(Y, Z=0) = -(Y - ground_depth)*sin(p), z = (depth - z_rad)/cos(p).
+    double p = pitch_deg * M_PI / 180.0;
+    double cos_p = std::cos(p), sin_p = std::sin(p);
+    double z_lo = 0.0, z_hi = 0.0;
+    bool first = true;
+    for (double y : {sar.y_min_m, sar.y_max_m}) {
+        for (double depth : {bp.min_depth, bp.max_depth}) {
+            double z_rad0 = -(y - sar.ground_depth_m) * sin_p;
+            double z = (depth - z_rad0) / cos_p;
+            if (first) { z_lo = z_hi = z; first = false; }
+            z_lo = std::min(z_lo, z);
+            z_hi = std::max(z_hi, z);
+        }
+    }
+    z_min = z_lo;
+    z_max = z_hi;
+}
+
 }  // namespace gpr

data_acq_and_processing/gpr_cpp/src/kernel_src.hpp

@@ -1,70 +1,81 @@
-// Исходник CUDA-ядра для NVRTC (порт bp_aperture_contribution из gpr_core/kernels.py).
-// Self-contained complex<float> (без cupy-заголовков).
+// Исходник CUDA-ядра для NVRTC (порт sar_frame_contribution из
+// Multi_antenna_3D_clean_v1.ipynb). Self-contained complex<float>.
+//
+// Геометрия (координаты, дальности, tau, вес компенсации) считается в double:
+// эталон ноутбука — float64, а в float32 позиция выборки A-скана (тысячи
+// отсчётов) теряет ~3e-4 отсчёта и расходится с эталоном на ~1e-4. A-сканы
+// остаются float32 — их значения O(1) и линейная интерполяция не теряет.
 #pragma once
 
 namespace gpr {
 
 inline const char* APERTURE_KERNEL_SRC = R"CUDA(
 struct cf { float x, y; };
-__device__ __forceinline__ cf operator+(cf a, cf b){ return {a.x+b.x, a.y+b.y}; }
-__device__ __forceinline__ cf operator-(cf a, cf b){ return {a.x-b.x, a.y-b.y}; }
-__device__ __forceinline__ cf operator*(float s, cf a){ return {s*a.x, s*a.y}; }
 
-// Весь взвешенный объём апертуры за один запуск: один поток на воксель,
-// внутренний цикл по n_contrib = n_frames*n_pairs. acc/cnt пишутся один раз.
+// Взвешенный объём апертуры: один поток на воксель, внутренний цикл по
+// n_contrib = n_frames*n_pairs. Обрабатывается диапазон вокселей
+// [gid_begin, gid_end) — хост запускает ядро чанками, переиспользуя небольшие
+// acc/cnt-буферы (полный объём в VRAM не обязан помещаться: на десктопе с
+// занятой другими процессами картой свободных может быть <200 МБ).
 extern "C" __global__ void bp_aperture_contribution(
     const cf* __restrict__ H,
-    const float t0, const float inv_dt, const int n_fft, const float inv_v,
-    const float* __restrict__ xt, const float* __restrict__ yt, const float* __restrict__ zt,
-    const float* __restrict__ xr, const float* __restrict__ yr, const float* __restrict__ zr,
-    const float* __restrict__ inv_geo_ref, const float* __restrict__ wgt,
-    const float* __restrict__ gx, const float* __restrict__ gy, const float* __restrict__ gz,
+    const double t0, const double inv_dt, const int n_fft, const double inv_v,
+    const double* __restrict__ xt, const double* __restrict__ yt, const double* __restrict__ zt,
+    const double* __restrict__ xr, const double* __restrict__ yr, const double* __restrict__ zr,
+    const double* __restrict__ inv_geo_ref, const float* __restrict__ wgt,
+    const double* __restrict__ gx, const double* __restrict__ gy, const double* __restrict__ gz,
     const int NX, const int NY, const int NZ, const int n_contrib,
-    const float cos_p, const float sin_p,
-    const float min_depth, const float max_depth,
-    const float comp_pow, const float w_max, const int use_comp,
+    const long gid_begin, const long gid_end,
+    const double cos_p, const double sin_p, const double ground_depth,
+    const double min_depth, const double max_depth,
+    const double comp_pow, const double w_max, const int use_comp,
     cf* __restrict__ acc, float* __restrict__ cnt)
 {
-    const long total = (long)NX * NY * NZ;
-    for (long gid = blockIdx.x * blockDim.x + threadIdx.x;
-         gid < total; gid += (long)gridDim.x * blockDim.x) {
+    for (long gid = gid_begin + blockIdx.x * blockDim.x + threadIdx.x;
+         gid < gid_end; gid += (long)gridDim.x * blockDim.x) {
         int ix = gid % NX;
         int iz = (gid / NX) % NZ;
         int iy = gid / ((long)NX * NZ);
-        float x = gx[ix];
-        float ys = gy[iy], zs = gz[iz];
-        float y_rad = ys * cos_p + zs * sin_p;
-        float z_rad = -ys * sin_p + zs * cos_p;
+        double x = gx[ix];
+        double zs = gz[iz];
+        // Сцена → оси радара: Y=0 сцены — земля, ground_depth ниже нижнего ряда
+        // антенн (порт sar_scene_to_radar_yz).
+        double y0 = gy[iy] - ground_depth;
+        double y_rad = y0 * cos_p + zs * sin_p;
+        double z_rad = -y0 * sin_p + zs * cos_p;
         bool depth_ok = (z_rad >= min_depth) && (z_rad <= max_depth);
 
-        cf a = {0.f, 0.f};
+        double ax = 0.0, ay = 0.0;
         float c = 0.f;
         for (int i = 0; i < n_contrib; ++i) {
-            float w = wgt[i];
-            float dxt = x - xt[i], dyt = y_rad - yt[i], dzt = z_rad - zt[i];
-            float dxr = x - xr[i], dyr = y_rad - yr[i], dzr = z_rad - zr[i];
-            float Rt = sqrtf(dxt*dxt + dyt*dyt + dzt*dzt);
-            float Rr = sqrtf(dxr*dxr + dyr*dyr + dzr*dzr);
-            float pos = ((Rt + Rr) * inv_v - t0) * inv_dt;
-            bool inside = (pos >= 0.f) && (pos <= (float)(n_fft - 1));
-            if (!inside) continue;
-            int k = (int)floorf(pos);
+            double dxt = x - xt[i], dyt = y_rad - yt[i], dzt = z_rad - zt[i];
+            double dxr = x - xr[i], dyr = y_rad - yr[i], dzr = z_rad - zr[i];
+            double Rt = sqrt(dxt*dxt + dyt*dyt + dzt*dzt);
+            double Rr = sqrt(dxr*dxr + dyr*dyr + dzr*dzr);
+            double pos = ((Rt + Rr) * inv_v - t0) * inv_dt;
+            bool inside = (pos >= 0.0) && (pos <= (double)(n_fft - 1));
+            if (!inside || !depth_ok) continue;
+            int k = (int)floor(pos);
             if (k < 0) k = 0;
             if (k > n_fft - 2) k = n_fft - 2;
-            float f = pos - (float)k;
+            float f = (float)(pos - (double)k);
             const cf* h = H + (long)i * n_fft;
-            cf hv = h[k] + f * (h[k+1] - h[k]);
+            float hx = h[k].x + f * (h[k+1].x - h[k].x);
+            float hy = h[k].y + f * (h[k+1].y - h[k].y);
+            double w = (double)wgt[i];
             if (use_comp) {
-                float ratio = (1.f/(Rt*Rr + 1e-12f) + 1e-30f) * inv_geo_ref[i];
-                float denom = (comp_pow == 0.25f) ? sqrtf(sqrtf(ratio)) : powf(ratio, comp_pow);
-                float ww = 1.f/(denom + 1e-12f);
-                ww = fminf(fmaxf(ww, 0.f), w_max);
-                hv = ww * hv;
+                double ratio = (1.0/(Rt*Rr + 1e-12) + 1e-30) * inv_geo_ref[i];
+                double ww = 1.0/(pow(ratio, comp_pow) + 1e-12);
+                ww = fmin(fmax(ww, 0.0), w_max);
+                w *= ww;
             }
-            if (depth_ok) { a = a + (w * hv); c += w; }
+            ax += w * (double)hx;
+            ay += w * (double)hy;
+            c += wgt[i];
         }
-        acc[gid] = a;
-        cnt[gid] = c;
+        acc[gid - gid_begin].x = (float)ax;
+        acc[gid - gid_begin].y = (float)ay;
+        cnt[gid - gid_begin] = c;
     }
 }
 )CUDA";

data_acq_and_processing/gpr_cpp/src/sar3d.cpp

@@ -31,11 +31,18 @@ VolumeRaw SarEngine::aperture_volume(const std::vector<ApertureFrame>& frames, c
     int NX = grid.NX(), NY = grid.NY(), NZ = grid.NZ();
     size_t nvox = (size_t)NX * NY * NZ;
 
+    double pitch = mot_.pitch_deg * M_PI / 180.0;
+    // Опорное расхождение пары — на плоскости земли на глубине comp_ref_depth.
+    double geo_ref_y = sar_.coord_mode_ground
+                           ? ground_plane_y(bp_.comp_ref_depth, sar_.ground_depth_m, pitch)
+                           : 0.0;
+
     // Плоские массивы по (кадр, пара)
     std::vector<c64> H_stack((size_t)n_contrib * n_fft);
-    std::vector<float> xt(n_contrib), yt(n_contrib), zt(n_contrib);
-    std::vector<float> xr(n_contrib), yr(n_contrib), zr(n_contrib);
-    std::vector<float> igr(n_contrib), wgt(n_contrib);
+    std::vector<double> xt(n_contrib), yt(n_contrib), zt(n_contrib);
+    std::vector<double> xr(n_contrib), yr(n_contrib), zr(n_contrib);
+    std::vector<double> igr(n_contrib);
+    std::vector<float> wgt(n_contrib);
     int idx = 0;
     for (int f = 0; f < n_frames; ++f) {
         const AScanSet& a = *frames[f].ascans;
@@ -49,62 +56,177 @@ VolumeRaw SarEngine::aperture_volume(const std::vector<ApertureFrame>& frames, c
             const c128* src = a.pair_ptr(p);
             c64* dst = H_stack.data() + (size_t)idx * n_fft;
             for (int k = 0; k < n_fft; ++k) dst[k] = c64((float)src[k].real(), (float)src[k].imag());
-            xt[idx] = (float)(geom_.x_tx[i] + dx); yt[idx] = (float)(geom_.y_tx[i] + dy);
-            zt[idx] = (float)(geom_.z_tx[i] + dz);
-            xr[idx] = (float)(geom_.x_rx[j] + dx); yr[idx] = (float)(geom_.y_rx[j] + dy);
-            zr[idx] = (float)(geom_.z_rx[j] + dz);
-            igr[idx] = (float)(1.0 / pair_geo_ref(geom_, bp_, i, j, geom_.bp_plane_y));
+            xt[idx] = geom_.x_tx[i] + dx; yt[idx] = geom_.y_tx[i] + dy;
+            zt[idx] = geom_.z_tx[i] + dz;
+            xr[idx] = geom_.x_rx[j] + dx; yr[idx] = geom_.y_rx[j] + dy;
+            zr[idx] = geom_.z_rx[j] + dz;
+            igr[idx] = 1.0 / pair_geo_ref(geom_, bp_, i, j, geo_ref_y);
             wgt[idx] = w;
             ++idx;
         }
     }
 
-    std::vector<float> gx(grid.x.begin(), grid.x.end());
-    std::vector<float> gy(grid.y.begin(), grid.y.end());
-    std::vector<float> gz(grid.z.begin(), grid.z.end());
-
-    float cos_p = 1.f, sin_p = 0.f;
+    double cos_p = 1.0, sin_p = 0.0, ground_depth = 0.0;
     if (sar_.coord_mode_ground) {
-        double p = mot_.pitch_deg * M_PI / 180.0;
-        cos_p = (float)std::cos(p); sin_p = (float)std::sin(p);
+        cos_p = std::cos(pitch);
+        sin_p = std::sin(pitch);
+        ground_depth = sar_.ground_depth_m;
     }
 
     // Заливка на устройство
     DeviceBuffer<c64> dH(H_stack.size()); dH.upload(H_stack.data(), H_stack.size());
-    DeviceBuffer<float> dxt(n_contrib), dyt(n_contrib), dzt(n_contrib);
-    DeviceBuffer<float> dxr(n_contrib), dyr(n_contrib), dzr(n_contrib);
-    DeviceBuffer<float> digr(n_contrib), dwgt(n_contrib);
+    DeviceBuffer<double> dxt(n_contrib), dyt(n_contrib), dzt(n_contrib);
+    DeviceBuffer<double> dxr(n_contrib), dyr(n_contrib), dzr(n_contrib);
+    DeviceBuffer<double> digr(n_contrib);
+    DeviceBuffer<float> dwgt(n_contrib);
     dxt.upload(xt.data(), n_contrib); dyt.upload(yt.data(), n_contrib); dzt.upload(zt.data(), n_contrib);
     dxr.upload(xr.data(), n_contrib); dyr.upload(yr.data(), n_contrib); dzr.upload(zr.data(), n_contrib);
     digr.upload(igr.data(), n_contrib); dwgt.upload(wgt.data(), n_contrib);
-    DeviceBuffer<float> dgx(NX), dgy(NY), dgz(NZ);
-    dgx.upload(gx.data(), NX); dgy.upload(gy.data(), NY); dgz.upload(gz.data(), NZ);
-    DeviceBuffer<c64> dacc(nvox);
-    DeviceBuffer<float> dcnt(nvox);
+    DeviceBuffer<double> dgx(NX), dgy(NY), dgz(NZ);
+    dgx.upload(grid.x.data(), NX); dgy.upload(grid.y.data(), NY); dgz.upload(grid.z.data(), NZ);
 
-    float f_t0 = (float)t0, f_inv_dt = (float)(1.0 / dt), f_inv_v = (float)(1.0 / geom_.v);
-    float min_d = (float)bp_.min_depth, max_d = (float)bp_.max_depth;
-    float comp_pow = (float)bp_.comp_range_power, w_max = (float)bp_.comp_weight_max;
+    // Объём считается чанками с переиспользуемыми acc/cnt-буферами: полные
+    // буферы (~115 МБ на сетку 9.6 млн вокселей) могут не влезать в остаток
+    // VRAM, когда карту делят десктоп-процессы.
+    const size_t chunk_vox = std::min<size_t>(nvox, 2'000'000);
+    DeviceBuffer<c64> dacc(chunk_vox);
+    DeviceBuffer<float> dcnt(chunk_vox);
+
+    double inv_dt = 1.0 / dt, inv_v = 1.0 / geom_.v;
+    double min_d = bp_.min_depth, max_d = bp_.max_depth;
+    double comp_pow = bp_.comp_range_power, w_max = bp_.comp_weight_max;
     int use_comp = sar_.use_compensation ? 1 : 0;
 
-    void* args[] = {
-        dH.addr(), &f_t0, &f_inv_dt, &n_fft, &f_inv_v,
-        dxt.addr(), dyt.addr(), dzt.addr(), dxr.addr(), dyr.addr(), dzr.addr(),
-        digr.addr(), dwgt.addr(), dgx.addr(), dgy.addr(), dgz.addr(),
-        &NX, &NY, &NZ, &n_contrib, &cos_p, &sin_p, &min_d, &max_d,
-        &comp_pow, &w_max, &use_comp, dacc.addr(), dcnt.addr()};
-
-    int block = 256;
-    int gridDim = (int)std::min<long>(65535, ((long)nvox + block - 1) / block);
-    cu_check(cuLaunchKernel(kernel_->func(), gridDim, 1, 1, block, 1, 1, 0, 0, args, nullptr),
-             "launch aperture");
-    cu_check(cuCtxSynchronize(), "sync");
-
     VolumeRaw out;
     out.NX = NX; out.NY = NY; out.NZ = NZ;
     out.acc.resize(nvox); out.cnt.resize(nvox);
-    dacc.download(out.acc.data(), nvox);
-    dcnt.download(out.cnt.data(), nvox);
+    int block = 256;
+    for (size_t begin = 0; begin < nvox; begin += chunk_vox) {
+        const size_t end = std::min(nvox, begin + chunk_vox);
+        long gid_begin = (long)begin, gid_end = (long)end;
+        void* args[] = {
+            dH.addr(), &t0, &inv_dt, &n_fft, &inv_v,
+            dxt.addr(), dyt.addr(), dzt.addr(), dxr.addr(), dyr.addr(), dzr.addr(),
+            digr.addr(), dwgt.addr(), dgx.addr(), dgy.addr(), dgz.addr(),
+            &NX, &NY, &NZ, &n_contrib, &gid_begin, &gid_end,
+            &cos_p, &sin_p, &ground_depth, &min_d, &max_d,
+            &comp_pow, &w_max, &use_comp, dacc.addr(), dcnt.addr()};
+        int gridDim = (int)std::min<long>(65535, ((long)(end - begin) + block - 1) / block);
+        cu_check(cuLaunchKernel(kernel_->func(), gridDim, 1, 1, block, 1, 1, 0, 0, args, nullptr),
+                 "launch aperture");
+        cu_check(cuCtxSynchronize(), "sync");
+        dacc.download(out.acc.data() + begin, end - begin);
+        dcnt.download(out.cnt.data() + begin, end - begin);
+    }
+    return out;
+}
+
+namespace {
+
+// scipy.ndimage.gaussian_filter: ядро радиуса int(truncate*sigma+0.5), reflect.
+std::vector<double> gaussian_kernel_1d(double sigma, int radius) {
+    std::vector<double> w(2 * radius + 1);
+    double sum = 0.0;
+    for (int i = -radius; i <= radius; ++i) {
+        w[i + radius] = std::exp(-0.5 * i * i / (sigma * sigma));
+        sum += w[i + radius];
+    }
+    for (auto& v : w) v /= sum;
+    return w;
+}
+
+// Корреляция вдоль оси с reflect-границами ('d c b a | a b c d').
+inline int reflect_index(int i, int n) {
+    while (i < 0 || i >= n) {
+        if (i < 0) i = -i - 1;
+        if (i >= n) i = 2 * n - i - 1;
+    }
+    return i;
+}
+
+}  // namespace
+
+FinishedVolume finish_volume(const VolumeRaw& raw, const Grid& grid, double pitch_deg,
+                             const BPConfig& bp, const SARConfig& sar) {
+    int NX = raw.NX, NY = raw.NY, NZ = raw.NZ;
+    size_t nvox = (size_t)NX * NY * NZ;
+    FinishedVolume out;
+    out.NX = NX; out.NY = NY; out.NZ = NZ;
+
+    std::vector<double> vol(nvox);
+    for (size_t i = 0; i < nvox; ++i)
+        vol[i] = std::abs(c128(raw.acc[i].real(), raw.acc[i].imag())) / ((double)raw.cnt[i] + 1e-12);
+
+    // Маска глубины в радарных осях (в ядре невалидные вкладов не получали,
+    // но |0/(0+eps)| уже 0; обнуление здесь повторяет ноутбук буквально).
+    if (sar.coord_mode_ground) {
+        double p = pitch_deg * M_PI / 180.0;
+        double cos_p = std::cos(p), sin_p = std::sin(p);
+        for (int iy = 0; iy < NY; ++iy) {
+            double y0 = grid.y[iy] - sar.ground_depth_m;
+            for (int iz = 0; iz < NZ; ++iz) {
+                double z_rad = -y0 * sin_p + grid.z[iz] * cos_p;
+                if (z_rad >= bp.min_depth && z_rad <= bp.max_depth) continue;
+                double* row = vol.data() + ((size_t)iy * NZ + iz) * NX;
+                std::fill(row, row + NX, 0.0);
+            }
+        }
+    }
+
+    // Плоскости Y для нормировки: без верхних/нижних norm_edge_margin_m.
+    std::vector<bool> ref_plane(NY, true);
+    double margin = sar.norm_edge_margin_m;
+    double y_span = grid.y.back() - grid.y.front();
+    if (margin > 0.0 && y_span > 2.5 * margin) {
+        for (int iy = 0; iy < NY; ++iy)
+            ref_plane[iy] = grid.y[iy] >= grid.y.front() + margin &&
+                            grid.y[iy] <= grid.y.back() - margin;
+    }
+    auto ref_max = [&](const std::vector<double>& v) {
+        double m = 0.0;
+        for (int iy = 0; iy < NY; ++iy) {
+            if (!ref_plane[iy]) continue;
+            const double* plane = v.data() + (size_t)iy * NZ * NX;
+            for (size_t i = 0; i < (size_t)NZ * NX; ++i) m = std::max(m, plane[i]);
+        }
+        return m + 1e-12;
+    };
+
+    double level_raw = ref_max(vol);
+    for (auto& v : vol) v /= level_raw;
+
+    // Гаусс sigma=(0, s, s): сглаживание по Z и X внутри каждого Y-плана
+    // (порядок осей как у scipy: сначала ось 1, затем ось 2).
+    if (sar.smooth_sigma > 0.0) {
+        int radius = (int)(4.0 * sar.smooth_sigma + 0.5);
+        const auto kernel = gaussian_kernel_1d(sar.smooth_sigma, radius);
+        std::vector<double> tmp((size_t)NZ * NX);
+        for (int iy = 0; iy < NY; ++iy) {
+            double* plane = vol.data() + (size_t)iy * NZ * NX;
+            // вдоль Z (ось 1)
+            for (int iz = 0; iz < NZ; ++iz)
+                for (int ix = 0; ix < NX; ++ix) {
+                    double s = 0.0;
+                    for (int k = -radius; k <= radius; ++k)
+                        s += kernel[k + radius] * plane[(size_t)reflect_index(iz + k, NZ) * NX + ix];
+                    tmp[(size_t)iz * NX + ix] = s;
+                }
+            // вдоль X (ось 2)
+            for (int iz = 0; iz < NZ; ++iz)
+                for (int ix = 0; ix < NX; ++ix) {
+                    double s = 0.0;
+                    for (int k = -radius; k <= radius; ++k)
+                        s += kernel[k + radius] * tmp[(size_t)iz * NX + reflect_index(ix + k, NX)];
+                    plane[(size_t)iz * NX + ix] = s;
+                }
+        }
+    }
+
+    double level_smooth = ref_max(vol);
+    out.volume.resize(nvox);
+    for (size_t i = 0; i < nvox; ++i)
+        out.volume[i] = (float)std::max(vol[i] / level_smooth, 0.0);
+    out.level = level_raw * level_smooth;
     return out;
 }
 

data_acq_and_processing/processing/data_processor/include/processing_live_config.hpp

@@ -90,21 +90,47 @@ struct ProcessingLiveConfig {
     float gpr_visible_x_max_m = 2.0F;
     float gpr_visible_z_min_m = 0.0F;
     float gpr_visible_z_max_m = 14.0F;
-    // 3D SAR (sar3d processor) parameters; defaults mirror VNA_AlGo gpr_core.
-    // Frequency band, depth gate, antenna selection and motion (speed, look
-    // angle, direction, phase correction) are shared with the gpr_* fields.
-    std::uint32_t sar3d_frame_radius = 6U;
-    float sar3d_edge_frame_weight = 0.75F;
-    float sar3d_pitch_deg = 6.0F;
-    float sar3d_center_depth_m = 5.6F;
-    float sar3d_half_width_x_m = 1.8F;
-    float sar3d_half_width_y_m = 1.0F;
-    float sar3d_half_width_z_m = 1.8F;
-    std::uint32_t sar3d_grid_nx = 150U;
-    std::uint32_t sar3d_grid_ny = 120U;
-    std::uint32_t sar3d_grid_nz = 160U;
+    // 3D SAR (sar3d processor) parameters; defaults mirror the reference
+    // notebook Multi_antenna_3D_clean_v1.ipynb (cells 0/4). Frequency band,
+    // depth gate, antenna selection and motion (speed, look angle, direction,
+    // phase correction) are shared with the gpr_* fields.
+    std::uint32_t sar3d_frame_radius = 0U;
+    float sar3d_edge_frame_weight = 0.8F;
+    // Frame pitch for ground coordinates and inter-frame offsets. Override
+    // disabled -> the pitch follows gpr_look_angle_deg (notebook
+    // BP_GROUND_PITCH_DEG = None).
+    bool sar3d_pitch_override_enabled = false;
+    float sar3d_pitch_deg = 17.5F;
+    // Height of the LOWER antenna row above the ground; scene Y=0 is the ground.
+    float sar3d_ground_depth_m = 1.25F;
+    // Scene grid bounds (metres): X across, Y height above the ground, Z along
+    // the ground away from the radar. Voxels whose radar depth falls outside
+    // [gpr_min_depth, gpr_max_depth] are zeroed by the mask regardless.
+    // Doubles, not floats: the parity tool feeds the notebook-derived bounds
+    // through here and a float roundtrip shifts the grid by ~1e-7 m, flipping
+    // knife-edge voxels of the depth mask against the reference volumes.
+    double sar3d_x_min_m = -2.0;
+    double sar3d_x_max_m = 2.0;
+    double sar3d_y_min_m = -0.5;
+    double sar3d_y_max_m = 2.0;
+    double sar3d_z_min_m = 0.0;
+    double sar3d_z_max_m = 8.0;
+    float sar3d_voxel_x_m = 0.025F;
+    float sar3d_voxel_y_m = 0.02F;
+    float sar3d_voxel_z_m = 0.02F;
+    // Volume finalization: normalization ignores this much of the grid's Y
+    // edges (constant-range ridge), then Gaussian smoothing over (Z, X).
+    float sar3d_norm_edge_margin_m = 0.3F;
+    float sar3d_smooth_sigma = 1.5F;
+    // Geometric spreading compensation 1/(Rtx*Rrx)^power, capped.
+    float sar3d_comp_range_power = 0.3F;
+    // Fallback frame segmentation when per-trace capture stamps are absent.
+    std::uint32_t sar3d_tx_per_segment = 2U;
     bool sar3d_use_compensation = true;
     bool sar3d_normalize_pairs = true;
+    // Debug: when non-empty, the processor dumps every finished volume to this
+    // directory (raw float32 + json header). Used by the parity tool only.
+    std::string sar3d_volume_dump_dir{};
     // Bottom-view projection angle around the vertical Y axis, degrees:
     // 0 = front view (X horizontal), 90 = side view (Z horizontal),
     // intermediate values project along the rotated horizontal direction.

data_acq_and_processing/processing/data_processor/src/processing_live_config.cpp

@@ -392,10 +392,13 @@ void apply_legacy_gpr_algorithm_alias(ProcessingLiveConfig& config, const std::s
     for (const auto& [key, target] : {
              std::pair{"sar3d_edge_frame_weight", &config.sar3d_edge_frame_weight},
              std::pair{"sar3d_pitch_deg", &config.sar3d_pitch_deg},
-             std::pair{"sar3d_center_depth_m", &config.sar3d_center_depth_m},
-             std::pair{"sar3d_half_width_x_m", &config.sar3d_half_width_x_m},
-             std::pair{"sar3d_half_width_y_m", &config.sar3d_half_width_y_m},
-             std::pair{"sar3d_half_width_z_m", &config.sar3d_half_width_z_m},
+             std::pair{"sar3d_ground_depth_m", &config.sar3d_ground_depth_m},
+             std::pair{"sar3d_voxel_x_m", &config.sar3d_voxel_x_m},
+             std::pair{"sar3d_voxel_y_m", &config.sar3d_voxel_y_m},
+             std::pair{"sar3d_voxel_z_m", &config.sar3d_voxel_z_m},
+             std::pair{"sar3d_norm_edge_margin_m", &config.sar3d_norm_edge_margin_m},
+             std::pair{"sar3d_smooth_sigma", &config.sar3d_smooth_sigma},
+             std::pair{"sar3d_comp_range_power", &config.sar3d_comp_range_power},
              std::pair{"sar3d_bottom_view_deg", &config.sar3d_bottom_view_deg},
          }) {
         if (const auto found = root.find(key); found != root.end()) {
@@ -406,25 +409,37 @@ void apply_legacy_gpr_algorithm_alias(ProcessingLiveConfig& config, const std::s
         }
     }
     for (const auto& [key, target] : {
-             std::pair{"sar3d_grid_nx", &config.sar3d_grid_nx},
-             std::pair{"sar3d_grid_ny", &config.sar3d_grid_ny},
-             std::pair{"sar3d_grid_nz", &config.sar3d_grid_nz},
+             std::pair{"sar3d_x_min_m", &config.sar3d_x_min_m},
+             std::pair{"sar3d_x_max_m", &config.sar3d_x_max_m},
+             std::pair{"sar3d_y_min_m", &config.sar3d_y_min_m},
+             std::pair{"sar3d_y_max_m", &config.sar3d_y_max_m},
+             std::pair{"sar3d_z_min_m", &config.sar3d_z_min_m},
+             std::pair{"sar3d_z_max_m", &config.sar3d_z_max_m},
          }) {
         if (const auto found = root.find(key); found != root.end()) {
-            *target = parse_u32_number(*found, std::string("processing.") + key);
+            if (!found->is_number()) {
+                throw std::runtime_error(std::string("processing.") + key + " must be number");
+            }
+            *target = found->get<double>();
         }
     }
-    if (const auto found = root.find("sar3d_use_compensation"); found != root.end()) {
-        if (!found->is_boolean()) {
-            throw std::runtime_error("processing.sar3d_use_compensation must be bool");
+    if (const auto found = root.find("sar3d_tx_per_segment"); found != root.end()) {
+        config.sar3d_tx_per_segment = parse_u32_number(*found, "processing.sar3d_tx_per_segment");
+        if (config.sar3d_tx_per_segment == 0U) {
+            throw std::runtime_error("processing.sar3d_tx_per_segment must be >= 1");
         }
-        config.sar3d_use_compensation = found->get<bool>();
     }
-    if (const auto found = root.find("sar3d_normalize_pairs"); found != root.end()) {
-        if (!found->is_boolean()) {
-            throw std::runtime_error("processing.sar3d_normalize_pairs must be bool");
+    for (const auto& [key, target] : {
+             std::pair{"sar3d_pitch_override_enabled", &config.sar3d_pitch_override_enabled},
+             std::pair{"sar3d_use_compensation", &config.sar3d_use_compensation},
+             std::pair{"sar3d_normalize_pairs", &config.sar3d_normalize_pairs},
+         }) {
+        if (const auto found = root.find(key); found != root.end()) {
+            if (!found->is_boolean()) {
+                throw std::runtime_error(std::string("processing.") + key + " must be bool");
+            }
+            *target = found->get<bool>();
         }
-        config.sar3d_normalize_pairs = found->get<bool>();
     }
     if (const auto found = root.find("ignore_socket_speed"); found != root.end()) {
         if (!found->is_boolean()) {

data_acq_and_processing/processing/processors/src/sar3d_processor.cpp

@@ -4,7 +4,10 @@
 #include <bit>
 #include <cmath>
 #include <cstdint>
+#include <cstdio>
+#include <fstream>
 #include <iostream>
+#include <limits>
 #include <map>
 #include <optional>
 #include <stdexcept>
@@ -19,7 +22,6 @@ namespace radar::processing {
 namespace {
 
 constexpr std::size_t kAscanCacheLimit = 64U;
-constexpr double kVolumeEps = 1e-12;
 
 void hash_mix(std::size_t& seed, std::uint64_t value) {
     seed ^= value + 0x9e3779b97f4a7c15ULL + (seed << 6U) + (seed >> 2U);
@@ -130,6 +132,7 @@ struct Selection {
     bp.pair_normalize = live_config.sar3d_normalize_pairs;
     bp.pair_norm_depth_min = bp.min_depth;
     bp.pair_norm_depth_max = bp.max_depth;
+    bp.comp_range_power = static_cast<double>(live_config.sar3d_comp_range_power);
     return bp;
 }
 
@@ -137,9 +140,13 @@ struct Selection {
     gpr::MotionConfig motion{};
     motion.speed_m_s = static_cast<double>(live_config.gpr_speed_m_s);
     motion.look_angle_deg = static_cast<double>(live_config.gpr_look_angle_deg);
-    motion.pitch_deg = static_cast<double>(live_config.sar3d_pitch_deg);
+    // Notebook BP_GROUND_PITCH_DEG = None -> the frame pitch follows the look angle.
+    motion.pitch_deg = live_config.sar3d_pitch_override_enabled
+                           ? static_cast<double>(live_config.sar3d_pitch_deg)
+                           : static_cast<double>(live_config.gpr_look_angle_deg);
     motion.direction_sign = static_cast<double>(live_config.gpr_direction_sign);
     motion.apply_freq_phase_correction = live_config.gpr_apply_freq_phase_correction;
+    motion.tx_per_segment = static_cast<int>(std::max<std::uint32_t>(1U, live_config.sar3d_tx_per_segment));
     return motion;
 }
 
@@ -147,20 +154,38 @@ struct Selection {
     gpr::SARConfig sar{};
     sar.frame_radius = static_cast<int>(std::max<std::uint32_t>(0U, live_config.sar3d_frame_radius));
     sar.edge_frame_weight = static_cast<double>(live_config.sar3d_edge_frame_weight);
-    sar.half_width_m = {
-        static_cast<double>(live_config.sar3d_half_width_x_m),
-        static_cast<double>(live_config.sar3d_half_width_y_m),
-        static_cast<double>(live_config.sar3d_half_width_z_m),
-    };
-    sar.grid = {
-        static_cast<int>(std::max<std::uint32_t>(2U, live_config.sar3d_grid_nx)),
-        static_cast<int>(std::max<std::uint32_t>(2U, live_config.sar3d_grid_ny)),
-        static_cast<int>(std::max<std::uint32_t>(2U, live_config.sar3d_grid_nz)),
+    sar.ground_depth_m = static_cast<double>(live_config.sar3d_ground_depth_m);
+    sar.x_min_m = static_cast<double>(live_config.sar3d_x_min_m);
+    sar.x_max_m = static_cast<double>(live_config.sar3d_x_max_m);
+    sar.y_min_m = static_cast<double>(live_config.sar3d_y_min_m);
+    sar.y_max_m = static_cast<double>(live_config.sar3d_y_max_m);
+    sar.z_min_m = static_cast<double>(live_config.sar3d_z_min_m);
+    sar.z_max_m = static_cast<double>(live_config.sar3d_z_max_m);
+    sar.voxel_m = {
+        static_cast<double>(live_config.sar3d_voxel_x_m),
+        static_cast<double>(live_config.sar3d_voxel_y_m),
+        static_cast<double>(live_config.sar3d_voxel_z_m),
     };
+    sar.norm_edge_margin_m = static_cast<double>(live_config.sar3d_norm_edge_margin_m);
+    sar.smooth_sigma = static_cast<double>(live_config.sar3d_smooth_sigma);
     sar.use_compensation = live_config.sar3d_use_compensation;
     return sar;
 }
 
+// Hardware Tx order: first occurrence of each output in run_combos. Deliberately
+// ignores the live Tx selection so segment/slot assignment stays a hardware
+// property (notebook tx_order_full).
+[[nodiscard]] auto tx_order_full(const config::RunConfig& run_config) -> std::vector<int> {
+    std::vector<int> order{};
+    for (const auto& combo : run_config.run_combos) {
+        const int output = static_cast<int>(combo.output_pos);
+        if (std::find(order.begin(), order.end(), output) == order.end()) {
+            order.push_back(output);
+        }
+    }
+    return order;
+}
+
 [[nodiscard]] auto ascan_fingerprint(
     const Selection& selection,
     const gpr::GeometryConfig& geom,
@@ -184,6 +209,7 @@ struct Selection {
     hash_mix(seed, motion.look_angle_deg);
     hash_mix(seed, motion.direction_sign);
     hash_mix(seed, static_cast<std::uint64_t>(motion.apply_freq_phase_correction ? 1U : 0U));
+    hash_mix(seed, static_cast<std::uint64_t>(motion.tx_per_segment));
     return seed;
 }
 
@@ -215,6 +241,7 @@ struct Selection {
     hash_mix(seed, bp.comp_weight_max);
     hash_mix(seed, bp.comp_ref_depth);
     hash_mix(seed, motion.pitch_deg);
+    hash_mix(seed, sar.ground_depth_m);
     hash_mix(seed, static_cast<std::uint64_t>(sar.use_compensation ? 1U : 0U));
     return seed;
 }
@@ -228,14 +255,80 @@ struct Selection {
     return static_cast<double>(collection.monotonic_ns) * 1e-9;
 }
 
-// One Tx's interleaved sweep spans the whole capture window, so the per-Tx
-// sweep time is capture_span / N_tx (same derivation as the coherent BP path).
-[[nodiscard]] auto tx_sweep_time_s(const ipc::PreprocessedCollection& collection, int n_tx) -> double {
-    if (collection.capture_end_ns <= collection.capture_start_ns || n_tx <= 0) {
-        return 0.0;
+// Real frame timing from the per-trace capture stamps: traces measured in one
+// segment share one window on both producers, so the first window seen per Tx
+// reconstructs the segments exactly (notebook snapshot_acquisition_timing).
+// Invalid when any trace carries the zero "unmeasured" sentinel.
+struct SnapshotTiming {
+    gpr::TxWindows windows{};    // seconds relative to the first trace start
+    double acq_s = 0.0;
+    double center_abs_s = 0.0;   // absolute acquisition centre (frame time)
+    bool valid = false;
+};
+
+[[nodiscard]] auto snapshot_timing(const ipc::PreprocessedCollection& collection) -> SnapshotTiming {
+    SnapshotTiming timing{};
+    if (collection.traces.empty()) {
+        return timing;
     }
-    const double span_s = static_cast<double>(collection.capture_end_ns - collection.capture_start_ns) * 1e-9;
-    return span_s / static_cast<double>(n_tx);
+    std::map<int, std::pair<std::uint64_t, std::uint64_t>> by_tx{};
+    std::uint64_t first = std::numeric_limits<std::uint64_t>::max();
+    std::uint64_t last = 0U;
+    for (const auto& trace : collection.traces) {
+        if (trace.capture_end_ns <= trace.capture_start_ns) {
+            return timing;
+        }
+        by_tx.try_emplace(
+            static_cast<int>(trace.combo.output_pos),
+            std::pair{trace.capture_start_ns, trace.capture_end_ns}
+        );
+        first = std::min(first, trace.capture_start_ns);
+        last = std::max(last, trace.capture_end_ns);
+    }
+    for (const auto& [tx, window] : by_tx) {
+        timing.windows[tx] = {
+            static_cast<double>(window.first - first) * 1e-9,
+            static_cast<double>(window.second - first) * 1e-9,
+        };
+    }
+    timing.acq_s = static_cast<double>(last - first) * 1e-9;
+    timing.center_abs_s = 0.5 * (static_cast<double>(first) + static_cast<double>(last)) * 1e-9;
+    timing.valid = true;
+    return timing;
+}
+
+// Debug dump for the parity tool: raw float32 volume [iy][iz][ix] plus a json
+// header with the axes and level.
+void dump_volume(
+    const std::string& dir,
+    std::uint64_t collection_id,
+    const gpr::FinishedVolume& finished,
+    const gpr::Grid& grid
+) {
+    char name[64];
+    std::snprintf(name, sizeof(name), "/sar3d_volume_%llu",
+                  static_cast<unsigned long long>(collection_id));
+    const std::string base = dir + name;
+    std::ofstream bin(base + ".f32", std::ios::binary | std::ios::trunc);
+    bin.write(
+        reinterpret_cast<const char*>(finished.volume.data()),
+        static_cast<std::streamsize>(finished.volume.size() * sizeof(float))
+    );
+    std::ofstream meta(base + ".json", std::ios::trunc);
+    meta.precision(17);
+    meta << "{\"nx\": " << finished.NX << ", \"ny\": " << finished.NY
+         << ", \"nz\": " << finished.NZ << ", \"level\": " << finished.level;
+    const auto write_axis = [&meta](const char* key, const std::vector<double>& axis) {
+        meta << ", \"" << key << "\": [";
+        for (std::size_t index = 0U; index < axis.size(); ++index) {
+            meta << (index > 0U ? "," : "") << axis[index];
+        }
+        meta << "]";
+    };
+    write_axis("x_m", grid.x);
+    write_axis("y_m", grid.y);
+    write_axis("z_m", grid.z);
+    meta << "}\n";
 }
 
 [[nodiscard]] auto build_image_payload(
@@ -285,14 +378,26 @@ struct Sar3dProcessor::State {
         const Selection& selection,
         const gpr::GeometryConfig& geom,
         const gpr::BPConfig& bp,
-        gpr::MotionConfig motion
+        const gpr::MotionConfig& motion,
+        const std::vector<int>& tx_order
     ) -> const CacheEntry& {
         if (const auto found = ascan_cache.find(collection.collection_id); found != ascan_cache.end()) {
             return found->second;
         }
 
         CacheEntry entry{};
-        entry.time_s = collection_time_s(collection);
+        auto timing = snapshot_timing(collection);
+        // The segmented model needs a window for every selected Tx; fall back to
+        // the uniform split otherwise (notebook median model).
+        if (timing.valid) {
+            for (const auto output : geom.output_idx) {
+                if (timing.windows.find(output) == timing.windows.end()) {
+                    timing.valid = false;
+                    break;
+                }
+            }
+        }
+        entry.time_s = timing.valid ? timing.center_abs_s : collection_time_s(collection);
 
         std::map<gpr::PairKey, std::vector<gpr::c128>> s21{};
         std::map<gpr::PairKey, std::vector<double>> freq{};
@@ -317,9 +422,25 @@ struct Sar3dProcessor::State {
         }
 
         if (s21.size() == selection.inputs.size() * selection.outputs.size()) {
-            motion.tx_sweep_time_s = tx_sweep_time_s(collection, geom.n_tx);
-            const auto timestamps = gpr::compute_pair_timestamps(motion, geom);
+            gpr::TxWindows windows{};
+            double acq_s = 0.0;
+            if (timing.valid) {
+                windows = std::move(timing.windows);
+                acq_s = timing.acq_s;
+            } else {
+                if (collection.capture_end_ns > collection.capture_start_ns) {
+                    acq_s = static_cast<double>(collection.capture_end_ns - collection.capture_start_ns) * 1e-9;
+                }
+                auto order = tx_order;
+                for (const auto output : geom.output_idx) {
+                    if (std::find(order.begin(), order.end(), output) == order.end()) {
+                        order.push_back(output);
+                    }
+                }
+                windows = gpr::uniform_tx_windows(order, acq_s, motion.tx_per_segment);
+            }
             try {
+                const auto timestamps = gpr::compute_pair_timestamps(windows, acq_s, tx_order, geom, motion);
                 entry.ascans = std::make_shared<gpr::AScanSet>(gpr::preprocess_snapshot(
                     s21, freq, /*background=*/nullptr, motion, geom, bp, timestamps, bp.pair_normalize
                 ));
@@ -399,12 +520,14 @@ auto Sar3dProcessor::process_collection(
         }
         window.push_back(&collection);
 
+        const auto tx_order = tx_order_full(run_config);
         std::vector<std::shared_ptr<gpr::AScanSet>> ascan_sets{};
         std::vector<double> frame_times{};
         ascan_sets.reserve(window.size());
         frame_times.reserve(window.size());
         for (const auto* entry_collection : window) {
-            const auto& entry = state_->get_or_compute_ascans(*entry_collection, selection, geom, bp, motion);
+            const auto& entry =
+                state_->get_or_compute_ascans(*entry_collection, selection, geom, bp, motion, tx_order);
             if (entry.ascans == nullptr) {
                 continue;
             }
@@ -450,14 +573,17 @@ auto Sar3dProcessor::process_collection(
             state_->engine_fp = engine_fp;
         }
 
-        const auto center_scene =
-            gpr::point_to_scene(0.0, 0.0, static_cast<double>(live_config.sar3d_center_depth_m), motion, sar);
-        const gpr::Grid grid = gpr::build_grid_scene(center_scene, sar);
-        const gpr::VolumeRaw volume = state_->engine->aperture_volume(frames, grid);
+        const gpr::Grid grid = gpr::build_full_region_grid(sar);
+        const gpr::VolumeRaw raw = state_->engine->aperture_volume(frames, grid);
+        // Full notebook finalization: |acc/cnt|, depth gate, normalization
+        // without the Y-edge planes, Gaussian smoothing over (Z, X), clip.
+        const gpr::FinishedVolume finished =
+            gpr::finish_volume(raw, grid, motion.pitch_deg, bp, sar);
+
+        if (!live_config.sar3d_volume_dump_dir.empty()) {
+            dump_volume(live_config.sar3d_volume_dump_dir, collection.collection_id, finished, grid);
+        }
 
-        // Finalization (port of gpr_core sar_finish_volume, minus smoothing):
-        // |acc/(cnt+eps)|, ground-coordinate depth gate, normalization to peak.
-        //
         // Two projections match the legacy render conventions (X across,
         // Z range/"вдаль", Y up):
         // - top view `sar3d_proj_xz`: max over Y, axes X (horizontal) / Z;
@@ -466,12 +592,9 @@ auto Sar3dProcessor::process_collection(
         //   `sar3d_bottom_view_deg` around Y: 0° = front view (u = X, max
         //   along Z), 90° = side view (u = Z, max along X), intermediate
         //   angles project along the rotated perpendicular.
-        const int nx = volume.NX;
-        const int ny = volume.NY;
-        const int nz = volume.NZ;
-        const double pitch_rad = motion.pitch_deg * M_PI / 180.0;
-        const double cos_p = std::cos(pitch_rad);
-        const double sin_p = std::sin(pitch_rad);
+        const int nx = finished.NX;
+        const int ny = finished.NY;
+        const int nz = finished.NZ;
 
         const double theta_deg = std::clamp(static_cast<double>(live_config.sar3d_bottom_view_deg), 0.0, 90.0);
         const double theta_rad = theta_deg * M_PI / 180.0;
@@ -501,24 +624,18 @@ auto Sar3dProcessor::process_collection(
         const double u_start = u_axis.front();
         const double inv_du = nu > 1 ? static_cast<double>(nu - 1) / (u_axis.back() - u_start) : 0.0;
 
+        // The volume is already normalized to ~[0, 1] by finish_volume, so the
+        // projections are published as-is; the scalar payload carries `level`
+        // (the absolute scale the volume was divided by).
         std::vector<float> projection_xz(static_cast<std::size_t>(nz) * nx, 0.0F);
         std::vector<float> projection_bottom(static_cast<std::size_t>(ny) * nu, 0.0F);
-        float peak = 0.0F;
         for (int iy = 0; iy < ny; ++iy) {
             for (int iz = 0; iz < nz; ++iz) {
-                const double z_rad = -grid.y[iy] * sin_p + grid.z[iz] * cos_p;
-                const bool depth_ok = z_rad >= bp.min_depth && z_rad <= bp.max_depth;
-                if (!depth_ok) {
-                    continue;
-                }
                 const std::size_t row = (static_cast<std::size_t>(iy) * nz + iz) * nx;
                 const std::size_t bottom_row = static_cast<std::size_t>(iy) * nu;
                 const double z_sin_t = grid.z[iz] * sin_t;
                 for (int ix = 0; ix < nx; ++ix) {
-                    const std::size_t voxel = row + static_cast<std::size_t>(ix);
-                    const float value = std::abs(volume.acc[voxel]) /
-                        (volume.cnt[voxel] + static_cast<float>(kVolumeEps));
-                    peak = std::max(peak, value);
+                    const float value = finished.volume[row + static_cast<std::size_t>(ix)];
                     auto& xz_cell = projection_xz[static_cast<std::size_t>(iz) * nx + ix];
                     xz_cell = std::max(xz_cell, value);
                     int bin = ix;
@@ -534,14 +651,6 @@ auto Sar3dProcessor::process_collection(
                 }
             }
         }
-        if (peak > 0.0F) {
-            for (auto& value : projection_xz) {
-                value /= peak;
-            }
-            for (auto& value : projection_bottom) {
-                value /= peak;
-            }
-        }
 
         results.collection_payloads.push_back(
             build_image_payload("sar3d_proj_xz", grid.x, grid.z, std::move(projection_xz))
@@ -552,7 +661,7 @@ auto Sar3dProcessor::process_collection(
         ipc::ResultPayload peak_payload{};
         peak_payload.processing_name = "sar3d_peak";
         peak_payload.kind = ipc::ResultKind::ScalarF32;
-        peak_payload.scalar_value = peak;
+        peak_payload.scalar_value = static_cast<float>(finished.level);
         results.collection_payloads.push_back(std::move(peak_payload));
     } catch (const std::exception& exc) {
         // A CUDA/NVRTC or data error must not kill the data_processor loop:

data_acq_and_processing/processing/tools/sar3d_parity_main.cpp

@@ -1,12 +1,18 @@
-// Parity check for the sar3d processor against the VNA_AlGo reference path.
+// Parity check of the sar3d processor against the reference notebook
+// Multi_antenna_3D_clean_v1.ipynb.
 //
-// Loads the same .npy snapshot dataset the gpr_cpp `standalone_bp` example
-// uses, feeds it through Sar3dProcessor (combo mapping, timestamps, A-scan
-// cache, aperture, finalization) and prints the volume peak and the argmax of
-// both projections. Compare against `standalone_bp <dir> <center>` output:
-// the peak |acc/cnt| and the (X, Y, Z) of the maximum voxel must match.
+// Feeds a preprocessed snapshot of an extracted replay dataset through
+// Sar3dProcessor exactly the way the live pipeline would (geometry and combo
+// order from the dataset's config_profile.json, per-trace capture windows from
+// the snapshot's meta.json) and dumps the finished volume for comparison with
+// the notebook's *_volume.npz export:
 //
-// Usage: sar3d_parity <preprocessed_dir> <center_snapshot_name>
+//   sar3d_parity <dataset_root> <center_snapshot> [--dump-dir <dir>]
+//   python_app/scripts/sar3d_volume_compare.py <dir>/sar3d_volume_<id> <ref.npz>
+//
+// <dataset_root> contains config_profile.json and preprocessed/<snapshot>/.
+// Defaults mirror the parameters the reference volumes were built with
+// (f 2..6 GHz, depth 2.5..8 m, speed 1.45 m/s, look angle 17.5 deg).
 #include <algorithm>
 #include <cctype>
 #include <complex>
@@ -19,6 +25,10 @@
 #include <string>
 #include <vector>
 
+#include <nlohmann/json.hpp>
+
+#include "gpr/config.hpp"
+#include "gpr/geometry.hpp"
 #include "processing_live_config.hpp"
 #include "run_config.hpp"
 #include "sar3d_processor.hpp"
@@ -26,8 +36,7 @@
 
 namespace {
 
-// Minimal numpy .npy loader (v1.0, 1-D complex128/complex64/float64/float32),
-// mirroring gpr_cpp/examples/standalone_bp.cpp.
+// Minimal numpy .npy loader (v1.0, 1-D complex128/complex64/float64/float32).
 struct NpyArray {
     std::vector<double> f64{};
     std::vector<std::complex<double>> c128{};
@@ -77,43 +86,53 @@ NpyArray load_npy(const std::string& path) {
     return array;
 }
 
-// VNA_AlGo reference geometry (gpr_core/config.py) keyed by switch positions.
-auto build_run_config() -> radar::config::RunConfig {
-    radar::config::RunConfig config{};
-    config.gpr.relative_permittivity = 1.0F;
-    config.gpr.tx_geometry = {
-        {.output_pos = 0U, .x_m = 0.785F, .y_m = 0.0F, .z_m = 0.0F},
-        {.output_pos = 1U, .x_m = -0.785F, .y_m = 0.0F, .z_m = 0.0F},
-    };
-    config.gpr.rx_geometry = {
-        {.input_pos = 0U, .x_m = -0.42F, .y_m = 0.0F, .z_m = 0.0F},
-        {.input_pos = 1U, .x_m = 0.48F, .y_m = 0.0F, .z_m = 0.0F},
-        {.input_pos = 2U, .x_m = 0.18F, .y_m = 0.0F, .z_m = 0.0F},
-        {.input_pos = 3U, .x_m = -0.12F, .y_m = 0.0F, .z_m = 0.0F},
-    };
-    return config;
-}
-
-// Live config matching the VNA_AlGo defaults (BPConfig/MotionConfig/SARConfig).
-auto build_live_config() -> radar::processing::ProcessingLiveConfig {
+// Live config matching the reference notebook run (npz meta parameters).
+// The default program grid is a fixed window (X -2..2, Z 0..8), while the
+// notebook derives its bounds from the antenna span and the depth gate — so
+// the notebook bounds are recomputed here and passed explicitly.
+auto build_live_config(const radar::config::RunConfig& run_config)
+    -> radar::processing::ProcessingLiveConfig {
     radar::processing::ProcessingLiveConfig live{};
     live.processor_mode = "sar3d";
     live.gpr_start_freq_mhz = 2000.0F;
     live.gpr_stop_freq_mhz = 6000.0F;
-    live.gpr_min_depth_m = 3.2F;
-    live.gpr_max_depth_m = 10.0F;
-    live.gpr_speed_m_s = 0.854F;
-    live.gpr_look_angle_deg = 6.0F;
+    live.gpr_min_depth_m = 2.5F;
+    live.gpr_max_depth_m = 8.0F;
+    live.gpr_speed_m_s = 1.45F;
+    live.gpr_look_angle_deg = 17.5F;
     live.gpr_direction_sign = 1.0F;
     live.gpr_apply_freq_phase_correction = true;
+
+    gpr::GeometryConfig geom{};
+    geom.input_idx.clear();
+    geom.output_idx.clear();
+    geom.rx_positions.clear();
+    geom.tx_positions.clear();
+    for (const auto& entry : run_config.gpr.rx_geometry) {
+        geom.input_idx.push_back(static_cast<int>(entry.input_pos));
+        geom.rx_positions.push_back({entry.x_m, entry.y_m, entry.z_m});
+    }
+    for (const auto& entry : run_config.gpr.tx_geometry) {
+        geom.output_idx.push_back(static_cast<int>(entry.output_pos));
+        geom.tx_positions.push_back({entry.x_m, entry.y_m, entry.z_m});
+    }
+    geom.finalize();
+    gpr::BPConfig bp{};
+    bp.min_depth = live.gpr_min_depth_m;
+    bp.max_depth = live.gpr_max_depth_m;
+    gpr::SARConfig sar{};
+    double x_min = 0.0, x_max = 0.0, z_min = 0.0, z_max = 0.0;
+    gpr::notebook_grid_bounds(geom, bp, sar, live.gpr_look_angle_deg, /*x_margin_m=*/2.0,
+                              x_min, x_max, z_min, z_max);
+    live.sar3d_x_min_m = x_min;
+    live.sar3d_x_max_m = x_max;
+    live.sar3d_z_min_m = z_min;
+    live.sar3d_z_max_m = z_max;
     return live;
 }
 
-constexpr double kTxSweepTimeS = 0.079;
-constexpr int kTxCount = 2;
-constexpr int kRxCount = 4;
-
 auto load_collection(
+    const radar::config::RunConfig& run_config,
     const std::string& root,
     const std::string& snapshot,
     std::uint64_t collection_id
@@ -121,46 +140,73 @@ auto load_collection(
     radar::ipc::PreprocessedCollection collection{};
     collection.collection_id = collection_id;
 
-    // Frame time from the `_ns<...>` folder suffix, identical to standalone_bp.
-    const auto ns_pos = snapshot.rfind("_ns");
-    if (ns_pos == std::string::npos) {
-        throw std::runtime_error("snapshot has no _ns suffix: " + snapshot);
-    }
-    const auto mid_ns = static_cast<std::uint64_t>(std::stoull(snapshot.substr(ns_pos + 3)));
-    // The processor derives tx_sweep_time from capture_span / n_tx and the frame
-    // time from the capture midpoint, so set the window symmetric around mid_ns.
-    const auto span_ns = static_cast<std::uint64_t>(kTxCount * kTxSweepTimeS * 1e9);
-    collection.capture_start_ns = mid_ns - span_ns / 2U;
-    collection.capture_end_ns = mid_ns + span_ns / 2U;
-    collection.monotonic_ns = mid_ns;
-
-    for (int rx = 0; rx < kRxCount; ++rx) {
-        for (int tx = 0; tx < kTxCount; ++tx) {
-            char path[512];
-            std::snprintf(path, sizeof(path), "%s/%s/i%d_o%d_s21.npy", root.c_str(), snapshot.c_str(), rx, tx);
-            const auto s21 = load_npy(path);
-            std::snprintf(path, sizeof(path), "%s/%s/i%d_o%d_freq.npy", root.c_str(), snapshot.c_str(), rx, tx);
-            const auto freq = load_npy(path);
-
-            radar::ipc::SweepTraceBlock trace{};
-            trace.combo = radar::ipc::ComboKey{
-                .input_pos = static_cast<std::uint32_t>(rx),
-                .output_pos = static_cast<std::uint32_t>(tx),
-            };
-            trace.frequency_hz.assign(freq.f64.begin(), freq.f64.end());
-            trace.s21.reserve(s21.c128.size());
-            for (const auto& sample : s21.c128) {
-                trace.s21.push_back(radar::ipc::Complex32{
-                    .re = static_cast<float>(sample.real()),
-                    .im = static_cast<float>(sample.imag()),
-                });
+    // Per-trace capture windows from the snapshot's meta.json; the segmented
+    // timing model is reconstructed from them exactly like in the notebook.
+    std::map<std::pair<std::uint32_t, std::uint32_t>, std::pair<std::uint64_t, std::uint64_t>> windows{};
+    {
+        std::ifstream in(root + "/preprocessed/" + snapshot + "/meta.json");
+        if (in) {
+            nlohmann::json meta{};
+            in >> meta;
+            for (const auto& trace : meta.value("traces", nlohmann::json::array())) {
+                if (!trace.contains("capture_start_ns") || !trace.contains("capture_end_ns")) {
+                    continue;
+                }
+                windows[{trace.at("input").get<std::uint32_t>(), trace.at("output").get<std::uint32_t>()}] = {
+                    trace.at("capture_start_ns").get<std::uint64_t>(),
+                    trace.at("capture_end_ns").get<std::uint64_t>(),
+                };
             }
-            trace.s11.resize(trace.s21.size());
-            trace.capture_start_ns = collection.capture_start_ns;
-            trace.capture_end_ns = collection.capture_end_ns;
-            collection.traces.push_back(std::move(trace));
         }
     }
+
+    std::uint64_t first = 0U;
+    std::uint64_t last = 0U;
+    for (const auto& [key, window] : windows) {
+        if (first == 0U || window.first < first) {
+            first = window.first;
+        }
+        last = std::max(last, window.second);
+    }
+    if (first == 0U) {
+        // No meta labels: fall back to the `_ns<...>` folder suffix.
+        const auto ns_pos = snapshot.rfind("_ns");
+        if (ns_pos == std::string::npos) {
+            throw std::runtime_error("snapshot has no meta.json and no _ns suffix: " + snapshot);
+        }
+        last = static_cast<std::uint64_t>(std::stoull(snapshot.substr(ns_pos + 3)));
+        first = last;
+    }
+    collection.capture_start_ns = first;
+    collection.capture_end_ns = last;
+    collection.monotonic_ns = (first + last) / 2U;
+
+    for (const auto& combo : run_config.run_combos) {
+        char path[512];
+        std::snprintf(path, sizeof(path), "%s/preprocessed/%s/i%u_o%u_s21.npy",
+                      root.c_str(), snapshot.c_str(), combo.input_pos, combo.output_pos);
+        const auto s21 = load_npy(path);
+        std::snprintf(path, sizeof(path), "%s/preprocessed/%s/i%u_o%u_freq.npy",
+                      root.c_str(), snapshot.c_str(), combo.input_pos, combo.output_pos);
+        const auto freq = load_npy(path);
+
+        radar::ipc::SweepTraceBlock trace{};
+        trace.combo = combo;
+        trace.frequency_hz.assign(freq.f64.begin(), freq.f64.end());
+        trace.s21.reserve(s21.c128.size());
+        for (const auto& sample : s21.c128) {
+            trace.s21.push_back(radar::ipc::Complex32{
+                .re = static_cast<float>(sample.real()),
+                .im = static_cast<float>(sample.imag()),
+            });
+        }
+        trace.s11.resize(trace.s21.size());
+        if (const auto found = windows.find({combo.input_pos, combo.output_pos}); found != windows.end()) {
+            trace.capture_start_ns = found->second.first;
+            trace.capture_end_ns = found->second.second;
+        }
+        collection.traces.push_back(std::move(trace));
+    }
     return collection;
 }
 
@@ -195,14 +241,21 @@ void print_projection_argmax(
 
 int main(int argc, char** argv) {
     if (argc < 3) {
-        std::printf("usage: %s <preprocessed_dir> <center_snapshot>\n", argv[0]);
+        std::printf("usage: %s <dataset_root> <center_snapshot> [--dump-dir <dir>]\n", argv[0]);
         return 1;
     }
     const std::string root = argv[1];
     const std::string center = argv[2];
+    std::string dump_dir{};
+    for (int index = 3; index + 1 < argc; ++index) {
+        if (std::string{argv[index]} == "--dump-dir") {
+            dump_dir = argv[index + 1];
+        }
+    }
 
+    const std::string preprocessed = root + "/preprocessed";
     std::vector<std::string> snapshots{};
-    if (DIR* dir = opendir(root.c_str())) {
+    if (DIR* dir = opendir(preprocessed.c_str())) {
         for (dirent* entry = nullptr; (entry = readdir(dir)) != nullptr;) {
             const std::string name = entry->d_name;
             if (!name.empty() && (std::isdigit(name[0]) != 0)) {
@@ -214,12 +267,19 @@ int main(int argc, char** argv) {
     std::sort(snapshots.begin(), snapshots.end());
     const auto center_it = std::find(snapshots.begin(), snapshots.end(), center);
     if (center_it == snapshots.end()) {
-        std::printf("center %s not found\n", center.c_str());
+        std::printf("center %s not found in %s\n", center.c_str(), preprocessed.c_str());
         return 1;
     }
     const auto center_index = static_cast<std::size_t>(center_it - snapshots.begin());
 
-    const auto live = build_live_config();
+    const auto run_config = radar::config::load_run_config(root + "/config_profile.json");
+    if (run_config.gpr.tx_geometry.empty() || run_config.gpr.rx_geometry.empty()) {
+        std::printf("FAIL: config_profile.json has no gpr geometry\n");
+        return 1;
+    }
+
+    auto live = build_live_config(run_config);
+    live.sar3d_volume_dump_dir = dump_dir;
     const auto radius = static_cast<std::size_t>(live.sar3d_frame_radius);
     if (center_index < radius || center_index + radius >= snapshots.size()) {
         std::printf("not enough snapshots around center for radius %zu\n", radius);
@@ -228,14 +288,13 @@ int main(int argc, char** argv) {
 
     // The processor centers the aperture `radius` frames before the newest
     // collection, so feeding [center-r .. center+r] in order reproduces the
-    // symmetric aperture of standalone_bp around `center`.
+    // symmetric aperture around `center`.
     std::vector<radar::ipc::PreprocessedCollection> history{};
     for (std::size_t index = center_index - radius; index <= center_index + radius; ++index) {
-        history.push_back(load_collection(root, snapshots[index], index));
+        history.push_back(load_collection(run_config, root, snapshots[index], index));
     }
     std::printf("aperture: %zu frames around %s\n", history.size(), center.c_str());
 
-    const auto run_config = build_run_config();
     radar::processing::Sar3dProcessor processor{};
     radar::ipc::ResultCollection result{};
     for (std::size_t index = 0U; index < history.size(); ++index) {
@@ -253,7 +312,7 @@ int main(int argc, char** argv) {
     }
     for (const auto& payload : result.collection_payloads) {
         if (payload.kind == radar::ipc::ResultKind::ScalarF32) {
-            std::printf("%s: %.6e (compare with standalone_bp max |acc/cnt|)\n",
+            std::printf("%s (volume level): %.9e\n",
                         payload.processing_name.c_str(), static_cast<double>(payload.scalar_value));
         }
     }
@@ -269,6 +328,7 @@ int main(int argc, char** argv) {
         std::span<const radar::ipc::PreprocessedCollection>(history.data(), history.size() - 1U);
     for (const float angle_deg : {0.0F, 30.0F, 60.0F, 90.0F}) {
         auto live_angle = live;
+        live_angle.sar3d_volume_dump_dir.clear();
         live_angle.sar3d_bottom_view_deg = angle_deg;
         const auto angle_result =
             processor.process_collection(run_config, history.back(), last_previous, live_angle);
@@ -279,6 +339,11 @@ int main(int argc, char** argv) {
             }
         }
     }
+    if (!dump_dir.empty()) {
+        std::printf("volume dumped to %s/sar3d_volume_%llu.{f32,json}\n",
+                    dump_dir.c_str(),
+                    static_cast<unsigned long long>(history.back().collection_id));
+    }
     std::printf("OK: sar3d processor path completed\n");
     return 0;
 }

python_app/gui/controllers/app_window_config/live_processing_mixin.py

@@ -90,14 +90,22 @@ _WEB_LIVE_SCHEMA = [
     ("gpr_speed_m_s", "Motion", _GPR_LIKE_MODES, _dual("_gpr_speed_m_s", {"legacy_gpr": "_legacy_gpr_speed_m_s", "sar3d": "_sar3d_speed_m_s"})),
     ("sar3d_frame_radius", "3D SAR", ("sar3d",), _attr("_sar3d_frame_radius")),
     ("sar3d_edge_frame_weight", "3D SAR", ("sar3d",), _attr("_sar3d_edge_frame_weight")),
+    ("sar3d_pitch_override_enabled", "3D SAR", ("sar3d",), _attr("_sar3d_pitch_override_enabled")),
     ("sar3d_pitch_deg", "3D SAR", ("sar3d",), _attr("_sar3d_pitch_deg")),
-    ("sar3d_center_depth_m", "3D SAR", ("sar3d",), _attr("_sar3d_center_depth_m")),
-    ("sar3d_half_width_x_m", "3D SAR", ("sar3d",), _attr("_sar3d_half_width_x_m")),
-    ("sar3d_half_width_y_m", "3D SAR", ("sar3d",), _attr("_sar3d_half_width_y_m")),
-    ("sar3d_half_width_z_m", "3D SAR", ("sar3d",), _attr("_sar3d_half_width_z_m")),
-    ("sar3d_grid_nx", "3D SAR", ("sar3d",), _attr("_sar3d_grid_nx")),
-    ("sar3d_grid_ny", "3D SAR", ("sar3d",), _attr("_sar3d_grid_ny")),
-    ("sar3d_grid_nz", "3D SAR", ("sar3d",), _attr("_sar3d_grid_nz")),
+    ("sar3d_ground_depth_m", "3D SAR", ("sar3d",), _attr("_sar3d_ground_depth_m")),
+    ("sar3d_x_min_m", "3D SAR", ("sar3d",), _attr("_sar3d_x_min_m")),
+    ("sar3d_x_max_m", "3D SAR", ("sar3d",), _attr("_sar3d_x_max_m")),
+    ("sar3d_y_min_m", "3D SAR", ("sar3d",), _attr("_sar3d_y_min_m")),
+    ("sar3d_y_max_m", "3D SAR", ("sar3d",), _attr("_sar3d_y_max_m")),
+    ("sar3d_z_min_m", "3D SAR", ("sar3d",), _attr("_sar3d_z_min_m")),
+    ("sar3d_z_max_m", "3D SAR", ("sar3d",), _attr("_sar3d_z_max_m")),
+    ("sar3d_voxel_x_m", "3D SAR", ("sar3d",), _attr("_sar3d_voxel_x_m")),
+    ("sar3d_voxel_y_m", "3D SAR", ("sar3d",), _attr("_sar3d_voxel_y_m")),
+    ("sar3d_voxel_z_m", "3D SAR", ("sar3d",), _attr("_sar3d_voxel_z_m")),
+    ("sar3d_norm_edge_margin_m", "3D SAR", ("sar3d",), _attr("_sar3d_norm_edge_margin_m")),
+    ("sar3d_smooth_sigma", "3D SAR", ("sar3d",), _attr("_sar3d_smooth_sigma")),
+    ("sar3d_comp_range_power", "3D SAR", ("sar3d",), _attr("_sar3d_comp_range_power")),
+    ("sar3d_tx_per_segment", "3D SAR", ("sar3d",), _attr("_sar3d_tx_per_segment")),
     ("sar3d_use_compensation", "3D SAR", ("sar3d",), _attr("_sar3d_use_compensation")),
     ("sar3d_normalize_pairs", "3D SAR", ("sar3d",), _attr("_sar3d_normalize_pairs")),
     ("sar3d_bottom_view_deg", "3D SAR", ("sar3d",), _attr("_sar3d_bottom_view")),
@@ -553,9 +561,8 @@ class AppWindowLiveProcessingMixin:
                 f"depth={self._sar3d_min_depth_m.value():g}..{self._sar3d_max_depth_m.value():g} m, "
                 f"freq={self._sar3d_start_freq_mhz.value():g}..{self._sar3d_stop_freq_mhz.value():g} MHz, "
                 f"frame_radius={self._sar3d_frame_radius.value()}, "
-                f"grid={self._sar3d_grid_nx.value()}x{self._sar3d_grid_ny.value()}x{self._sar3d_grid_nz.value()}, "
-                f"center_depth={self._sar3d_center_depth_m.value():g} m, "
-                f"pitch={self._sar3d_pitch_deg.value():g} deg)"
+                f"voxel={self._sar3d_voxel_x_m.value():g}x{self._sar3d_voxel_y_m.value():g}x{self._sar3d_voxel_z_m.value():g} m, "
+                f"ground_depth={self._sar3d_ground_depth_m.value():g} m)"
             )
 
     def _clear_history_mode_caches(self) -> None:

python_app/gui/controllers/app_window_config/profile_io_mixin.py

@@ -350,14 +350,22 @@ class AppWindowConfigProfileIOMixin:
             self._sar3d_apply_freq_phase_correction,
             self._sar3d_frame_radius,
             self._sar3d_edge_frame_weight,
+            self._sar3d_pitch_override_enabled,
             self._sar3d_pitch_deg,
-            self._sar3d_center_depth_m,
-            self._sar3d_half_width_x_m,
-            self._sar3d_half_width_y_m,
-            self._sar3d_half_width_z_m,
-            self._sar3d_grid_nx,
-            self._sar3d_grid_ny,
-            self._sar3d_grid_nz,
+            self._sar3d_ground_depth_m,
+            self._sar3d_x_min_m,
+            self._sar3d_x_max_m,
+            self._sar3d_y_min_m,
+            self._sar3d_y_max_m,
+            self._sar3d_z_min_m,
+            self._sar3d_z_max_m,
+            self._sar3d_voxel_x_m,
+            self._sar3d_voxel_y_m,
+            self._sar3d_voxel_z_m,
+            self._sar3d_norm_edge_margin_m,
+            self._sar3d_smooth_sigma,
+            self._sar3d_comp_range_power,
+            self._sar3d_tx_per_segment,
             self._sar3d_use_compensation,
             self._sar3d_normalize_pairs,
             self._save_count,
@@ -562,14 +570,22 @@ class AppWindowConfigProfileIOMixin:
             self._sar3d_apply_freq_phase_correction.setChecked(bool(sar3d_state.apply_freq_phase_correction))
             self._sar3d_frame_radius.setValue(int(sar3d_state.frame_radius))
             self._sar3d_edge_frame_weight.setValue(float(sar3d_state.edge_frame_weight))
+            self._sar3d_pitch_override_enabled.setChecked(bool(sar3d_state.pitch_override_enabled))
             self._sar3d_pitch_deg.setValue(float(sar3d_state.pitch_deg))
-            self._sar3d_center_depth_m.setValue(float(sar3d_state.center_depth_m))
-            self._sar3d_half_width_x_m.setValue(float(sar3d_state.half_width_x_m))
-            self._sar3d_half_width_y_m.setValue(float(sar3d_state.half_width_y_m))
-            self._sar3d_half_width_z_m.setValue(float(sar3d_state.half_width_z_m))
-            self._sar3d_grid_nx.setValue(int(sar3d_state.grid_nx))
-            self._sar3d_grid_ny.setValue(int(sar3d_state.grid_ny))
-            self._sar3d_grid_nz.setValue(int(sar3d_state.grid_nz))
+            self._sar3d_ground_depth_m.setValue(float(sar3d_state.ground_depth_m))
+            self._sar3d_x_min_m.setValue(float(sar3d_state.x_min_m))
+            self._sar3d_x_max_m.setValue(float(sar3d_state.x_max_m))
+            self._sar3d_y_min_m.setValue(float(sar3d_state.y_min_m))
+            self._sar3d_y_max_m.setValue(float(sar3d_state.y_max_m))
+            self._sar3d_z_min_m.setValue(float(sar3d_state.z_min_m))
+            self._sar3d_z_max_m.setValue(float(sar3d_state.z_max_m))
+            self._sar3d_voxel_x_m.setValue(float(sar3d_state.voxel_x_m))
+            self._sar3d_voxel_y_m.setValue(float(sar3d_state.voxel_y_m))
+            self._sar3d_voxel_z_m.setValue(float(sar3d_state.voxel_z_m))
+            self._sar3d_norm_edge_margin_m.setValue(float(sar3d_state.norm_edge_margin_m))
+            self._sar3d_smooth_sigma.setValue(float(sar3d_state.smooth_sigma))
+            self._sar3d_comp_range_power.setValue(float(sar3d_state.comp_range_power))
+            self._sar3d_tx_per_segment.setValue(int(sar3d_state.tx_per_segment))
             self._sar3d_use_compensation.setChecked(bool(sar3d_state.use_compensation))
             self._sar3d_normalize_pairs.setChecked(bool(sar3d_state.normalize_pairs))
             self._set_combo_current_text(

python_app/gui/controllers/app_window_config/state_builders.py

@@ -438,14 +438,22 @@ class AppWindowConfigStateBuildersMixin:
                     apply_freq_phase_correction=bool(self._sar3d_apply_freq_phase_correction.isChecked()),
                     frame_radius=int(self._sar3d_frame_radius.value()),
                     edge_frame_weight=float(self._sar3d_edge_frame_weight.value()),
+                    pitch_override_enabled=bool(self._sar3d_pitch_override_enabled.isChecked()),
                     pitch_deg=float(self._sar3d_pitch_deg.value()),
-                    center_depth_m=float(self._sar3d_center_depth_m.value()),
-                    half_width_x_m=float(self._sar3d_half_width_x_m.value()),
-                    half_width_y_m=float(self._sar3d_half_width_y_m.value()),
-                    half_width_z_m=float(self._sar3d_half_width_z_m.value()),
-                    grid_nx=int(self._sar3d_grid_nx.value()),
-                    grid_ny=int(self._sar3d_grid_ny.value()),
-                    grid_nz=int(self._sar3d_grid_nz.value()),
+                    ground_depth_m=float(self._sar3d_ground_depth_m.value()),
+                    x_min_m=float(self._sar3d_x_min_m.value()),
+                    x_max_m=float(self._sar3d_x_max_m.value()),
+                    y_min_m=float(self._sar3d_y_min_m.value()),
+                    y_max_m=float(self._sar3d_y_max_m.value()),
+                    z_min_m=float(self._sar3d_z_min_m.value()),
+                    z_max_m=float(self._sar3d_z_max_m.value()),
+                    voxel_x_m=float(self._sar3d_voxel_x_m.value()),
+                    voxel_y_m=float(self._sar3d_voxel_y_m.value()),
+                    voxel_z_m=float(self._sar3d_voxel_z_m.value()),
+                    norm_edge_margin_m=float(self._sar3d_norm_edge_margin_m.value()),
+                    smooth_sigma=float(self._sar3d_smooth_sigma.value()),
+                    comp_range_power=float(self._sar3d_comp_range_power.value()),
+                    tx_per_segment=int(self._sar3d_tx_per_segment.value()),
                     use_compensation=bool(self._sar3d_use_compensation.isChecked()),
                     normalize_pairs=bool(self._sar3d_normalize_pairs.isChecked()),
                     bottom_view_deg=sar3d_bottom_view_deg_from_label(

python_app/gui/controllers/sections/processing_section.py

@@ -660,49 +660,114 @@ def build_processing_group(owner) -> QGroupBox:
         "Weight of the outermost aperture frames (1 = flat window, 0 = strong taper)."
     )
 
+    owner._sar3d_pitch_override_enabled = QCheckBox("Override pitch")
+    owner._sar3d_pitch_override_enabled.setChecked(bool(sar3d_defaults.pitch_override_enabled))
+    owner._sar3d_pitch_override_enabled.setToolTip(
+        "When unchecked, the frame pitch follows the look angle (notebook "
+        "BP_GROUND_PITCH_DEG = None); check to set it separately below."
+    )
+
     owner._sar3d_pitch_deg = QDoubleSpinBox()
     owner._sar3d_pitch_deg.setDecimals(2)
     owner._sar3d_pitch_deg.setRange(-90.0, 90.0)
     owner._sar3d_pitch_deg.setSingleStep(0.5)
     owner._sar3d_pitch_deg.setValue(float(sar3d_defaults.pitch_deg))
-    owner._sar3d_pitch_deg.setToolTip("Antenna pitch above the horizon (deg); tilts the imaged volume.")
+    owner._sar3d_pitch_deg.setToolTip(
+        "Frame pitch (deg) used for ground coordinates and aperture offsets when the override is on."
+    )
 
-    owner._sar3d_center_depth_m = QDoubleSpinBox()
-    owner._sar3d_center_depth_m.setDecimals(2)
-    owner._sar3d_center_depth_m.setRange(0.0, 50.0)
-    owner._sar3d_center_depth_m.setSingleStep(0.1)
-    owner._sar3d_center_depth_m.setValue(float(sar3d_defaults.center_depth_m))
-    owner._sar3d_center_depth_m.setToolTip("Range distance to the center of the imaged volume (m).")
+    owner._sar3d_ground_depth_m = QDoubleSpinBox()
+    owner._sar3d_ground_depth_m.setDecimals(2)
+    owner._sar3d_ground_depth_m.setRange(0.0, 10.0)
+    owner._sar3d_ground_depth_m.setSingleStep(0.05)
+    owner._sar3d_ground_depth_m.setValue(float(sar3d_defaults.ground_depth_m))
+    owner._sar3d_ground_depth_m.setToolTip(
+        "Height of the lower antenna row above the ground (m); scene Y=0 is the ground."
+    )
 
-    owner._sar3d_half_width_x_m = QDoubleSpinBox()
-    owner._sar3d_half_width_x_m.setDecimals(2)
-    owner._sar3d_half_width_x_m.setRange(0.01, 50.0)
-    owner._sar3d_half_width_x_m.setSingleStep(0.1)
-    owner._sar3d_half_width_x_m.setValue(float(sar3d_defaults.half_width_x_m))
+    owner._sar3d_x_min_m = QDoubleSpinBox()
+    owner._sar3d_x_min_m.setDecimals(2)
+    owner._sar3d_x_min_m.setRange(-50.0, 50.0)
+    owner._sar3d_x_min_m.setSingleStep(0.1)
+    owner._sar3d_x_min_m.setValue(float(sar3d_defaults.x_min_m))
 
-    owner._sar3d_half_width_y_m = QDoubleSpinBox()
-    owner._sar3d_half_width_y_m.setDecimals(2)
-    owner._sar3d_half_width_y_m.setRange(0.01, 50.0)
-    owner._sar3d_half_width_y_m.setSingleStep(0.1)
-    owner._sar3d_half_width_y_m.setValue(float(sar3d_defaults.half_width_y_m))
+    owner._sar3d_x_max_m = QDoubleSpinBox()
+    owner._sar3d_x_max_m.setDecimals(2)
+    owner._sar3d_x_max_m.setRange(-50.0, 50.0)
+    owner._sar3d_x_max_m.setSingleStep(0.1)
+    owner._sar3d_x_max_m.setValue(float(sar3d_defaults.x_max_m))
 
-    owner._sar3d_half_width_z_m = QDoubleSpinBox()
-    owner._sar3d_half_width_z_m.setDecimals(2)
-    owner._sar3d_half_width_z_m.setRange(0.01, 50.0)
-    owner._sar3d_half_width_z_m.setSingleStep(0.1)
-    owner._sar3d_half_width_z_m.setValue(float(sar3d_defaults.half_width_z_m))
+    owner._sar3d_z_min_m = QDoubleSpinBox()
+    owner._sar3d_z_min_m.setDecimals(2)
+    owner._sar3d_z_min_m.setRange(-50.0, 50.0)
+    owner._sar3d_z_min_m.setSingleStep(0.1)
+    owner._sar3d_z_min_m.setValue(float(sar3d_defaults.z_min_m))
 
-    owner._sar3d_grid_nx = QSpinBox()
-    owner._sar3d_grid_nx.setRange(2, 1024)
-    owner._sar3d_grid_nx.setValue(int(sar3d_defaults.grid_nx))
+    owner._sar3d_z_max_m = QDoubleSpinBox()
+    owner._sar3d_z_max_m.setDecimals(2)
+    owner._sar3d_z_max_m.setRange(-50.0, 50.0)
+    owner._sar3d_z_max_m.setSingleStep(0.1)
+    owner._sar3d_z_max_m.setValue(float(sar3d_defaults.z_max_m))
 
-    owner._sar3d_grid_ny = QSpinBox()
-    owner._sar3d_grid_ny.setRange(2, 1024)
-    owner._sar3d_grid_ny.setValue(int(sar3d_defaults.grid_ny))
+    owner._sar3d_y_min_m = QDoubleSpinBox()
+    owner._sar3d_y_min_m.setDecimals(2)
+    owner._sar3d_y_min_m.setRange(-10.0, 10.0)
+    owner._sar3d_y_min_m.setSingleStep(0.1)
+    owner._sar3d_y_min_m.setValue(float(sar3d_defaults.y_min_m))
 
-    owner._sar3d_grid_nz = QSpinBox()
-    owner._sar3d_grid_nz.setRange(2, 1024)
-    owner._sar3d_grid_nz.setValue(int(sar3d_defaults.grid_nz))
+    owner._sar3d_y_max_m = QDoubleSpinBox()
+    owner._sar3d_y_max_m.setDecimals(2)
+    owner._sar3d_y_max_m.setRange(-10.0, 10.0)
+    owner._sar3d_y_max_m.setSingleStep(0.1)
+    owner._sar3d_y_max_m.setValue(float(sar3d_defaults.y_max_m))
+
+    owner._sar3d_voxel_x_m = QDoubleSpinBox()
+    owner._sar3d_voxel_x_m.setDecimals(3)
+    owner._sar3d_voxel_x_m.setRange(0.005, 1.0)
+    owner._sar3d_voxel_x_m.setSingleStep(0.005)
+    owner._sar3d_voxel_x_m.setValue(float(sar3d_defaults.voxel_x_m))
+
+    owner._sar3d_voxel_y_m = QDoubleSpinBox()
+    owner._sar3d_voxel_y_m.setDecimals(3)
+    owner._sar3d_voxel_y_m.setRange(0.005, 1.0)
+    owner._sar3d_voxel_y_m.setSingleStep(0.005)
+    owner._sar3d_voxel_y_m.setValue(float(sar3d_defaults.voxel_y_m))
+
+    owner._sar3d_voxel_z_m = QDoubleSpinBox()
+    owner._sar3d_voxel_z_m.setDecimals(3)
+    owner._sar3d_voxel_z_m.setRange(0.005, 1.0)
+    owner._sar3d_voxel_z_m.setSingleStep(0.005)
+    owner._sar3d_voxel_z_m.setValue(float(sar3d_defaults.voxel_z_m))
+
+    owner._sar3d_norm_edge_margin_m = QDoubleSpinBox()
+    owner._sar3d_norm_edge_margin_m.setDecimals(2)
+    owner._sar3d_norm_edge_margin_m.setRange(0.0, 5.0)
+    owner._sar3d_norm_edge_margin_m.setSingleStep(0.05)
+    owner._sar3d_norm_edge_margin_m.setValue(float(sar3d_defaults.norm_edge_margin_m))
+    owner._sar3d_norm_edge_margin_m.setToolTip(
+        "Grid Y edges excluded from volume normalization (constant-range ridge)."
+    )
+
+    owner._sar3d_smooth_sigma = QDoubleSpinBox()
+    owner._sar3d_smooth_sigma.setDecimals(2)
+    owner._sar3d_smooth_sigma.setRange(0.0, 10.0)
+    owner._sar3d_smooth_sigma.setSingleStep(0.25)
+    owner._sar3d_smooth_sigma.setValue(float(sar3d_defaults.smooth_sigma))
+    owner._sar3d_smooth_sigma.setToolTip("Gaussian smoothing sigma over (Z, X) planes, voxels.")
+
+    owner._sar3d_comp_range_power = QDoubleSpinBox()
+    owner._sar3d_comp_range_power.setDecimals(2)
+    owner._sar3d_comp_range_power.setRange(0.0, 2.0)
+    owner._sar3d_comp_range_power.setSingleStep(0.05)
+    owner._sar3d_comp_range_power.setValue(float(sar3d_defaults.comp_range_power))
+    owner._sar3d_comp_range_power.setToolTip("Geometric spreading compensation power.")
+
+    owner._sar3d_tx_per_segment = QSpinBox()
+    owner._sar3d_tx_per_segment.setRange(1, 64)
+    owner._sar3d_tx_per_segment.setValue(int(sar3d_defaults.tx_per_segment))
+    owner._sar3d_tx_per_segment.setToolTip(
+        "Fallback Tx-per-segment split of a frame when per-trace capture stamps are absent."
+    )
 
     owner._sar3d_use_compensation = QCheckBox("Use per-voxel compensation")
     owner._sar3d_use_compensation.setChecked(bool(sar3d_defaults.use_compensation))
@@ -738,19 +803,27 @@ def build_processing_group(owner) -> QGroupBox:
             owner._sar3d_apply_freq_phase_correction,
             ("Frame radius", owner._sar3d_frame_radius),
             ("Edge frame weight", owner._sar3d_edge_frame_weight),
+            owner._sar3d_pitch_override_enabled,
             ("Pitch deg", owner._sar3d_pitch_deg),
-            ("Center depth m", owner._sar3d_center_depth_m),
-            ("Half width X m", owner._sar3d_half_width_x_m),
-            ("Half width Y m", owner._sar3d_half_width_y_m),
-            ("Half width Z m", owner._sar3d_half_width_z_m),
-            ("Grid Nx", owner._sar3d_grid_nx),
-            ("Grid Ny", owner._sar3d_grid_ny),
-            ("Grid Nz", owner._sar3d_grid_nz),
+            ("Ground depth m", owner._sar3d_ground_depth_m),
+            ("X min m", owner._sar3d_x_min_m),
+            ("X max m", owner._sar3d_x_max_m),
+            ("Y min m", owner._sar3d_y_min_m),
+            ("Y max m", owner._sar3d_y_max_m),
+            ("Z min m", owner._sar3d_z_min_m),
+            ("Z max m", owner._sar3d_z_max_m),
+            ("Voxel X m", owner._sar3d_voxel_x_m),
+            ("Voxel Y m", owner._sar3d_voxel_y_m),
+            ("Voxel Z m", owner._sar3d_voxel_z_m),
+            ("Norm edge margin m", owner._sar3d_norm_edge_margin_m),
+            ("Smooth sigma", owner._sar3d_smooth_sigma),
+            ("Comp range power", owner._sar3d_comp_range_power),
+            ("Tx per segment", owner._sar3d_tx_per_segment),
             ("Bottom view", owner._sar3d_bottom_view),
             owner._sar3d_use_compensation,
             owner._sar3d_normalize_pairs,
         ],
-        split_index=11,
+        split_index=13,
     )
     owner._processing_mode_pages.addWidget(sar3d_page)
 
@@ -839,14 +912,22 @@ def build_processing_group(owner) -> QGroupBox:
     owner._sar3d_apply_freq_phase_correction.toggled.connect(owner._on_processing_live_settings_changed)
     owner._sar3d_frame_radius.valueChanged.connect(owner._on_processing_live_settings_changed)
     owner._sar3d_edge_frame_weight.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_pitch_override_enabled.toggled.connect(owner._on_processing_live_settings_changed)
     owner._sar3d_pitch_deg.valueChanged.connect(owner._on_processing_live_settings_changed)
-    owner._sar3d_center_depth_m.valueChanged.connect(owner._on_processing_live_settings_changed)
-    owner._sar3d_half_width_x_m.valueChanged.connect(owner._on_processing_live_settings_changed)
-    owner._sar3d_half_width_y_m.valueChanged.connect(owner._on_processing_live_settings_changed)
-    owner._sar3d_half_width_z_m.valueChanged.connect(owner._on_processing_live_settings_changed)
-    owner._sar3d_grid_nx.valueChanged.connect(owner._on_processing_live_settings_changed)
-    owner._sar3d_grid_ny.valueChanged.connect(owner._on_processing_live_settings_changed)
-    owner._sar3d_grid_nz.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_ground_depth_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_x_min_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_x_max_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_z_min_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_z_max_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_y_min_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_y_max_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_voxel_x_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_voxel_y_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_voxel_z_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_norm_edge_margin_m.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_smooth_sigma.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_comp_range_power.valueChanged.connect(owner._on_processing_live_settings_changed)
+    owner._sar3d_tx_per_segment.valueChanged.connect(owner._on_processing_live_settings_changed)
     owner._sar3d_use_compensation.toggled.connect(owner._on_processing_live_settings_changed)
     owner._sar3d_normalize_pairs.toggled.connect(owner._on_processing_live_settings_changed)
     owner._sar3d_bottom_view.currentTextChanged.connect(owner._on_processing_live_settings_changed)

python_app/models/gui_profile_codec.py

@@ -533,39 +533,102 @@ def gui_profile_from_dict(payload: dict[str, Any]) -> GuiProfileModel:
                 gui.processing.sar3d.edge_frame_weight,
                 "gui.processing.sar3d",
             ),
+            pitch_override_enabled=_optional_bool(
+                sar3d_object,
+                "pitch_override_enabled",
+                gui.processing.sar3d.pitch_override_enabled,
+                "gui.processing.sar3d",
+            ),
             pitch_deg=_optional_float(
                 sar3d_object,
                 "pitch_deg",
                 gui.processing.sar3d.pitch_deg,
                 "gui.processing.sar3d",
             ),
-            center_depth_m=_optional_float(
+            ground_depth_m=_optional_float(
                 sar3d_object,
-                "center_depth_m",
-                gui.processing.sar3d.center_depth_m,
+                "ground_depth_m",
+                gui.processing.sar3d.ground_depth_m,
                 "gui.processing.sar3d",
             ),
-            half_width_x_m=_optional_float(
+            x_min_m=_optional_float(
                 sar3d_object,
-                "half_width_x_m",
-                gui.processing.sar3d.half_width_x_m,
+                "x_min_m",
+                gui.processing.sar3d.x_min_m,
                 "gui.processing.sar3d",
             ),
-            half_width_y_m=_optional_float(
+            x_max_m=_optional_float(
                 sar3d_object,
-                "half_width_y_m",
-                gui.processing.sar3d.half_width_y_m,
+                "x_max_m",
+                gui.processing.sar3d.x_max_m,
                 "gui.processing.sar3d",
             ),
-            half_width_z_m=_optional_float(
+            y_min_m=_optional_float(
                 sar3d_object,
-                "half_width_z_m",
-                gui.processing.sar3d.half_width_z_m,
+                "y_min_m",
+                gui.processing.sar3d.y_min_m,
+                "gui.processing.sar3d",
+            ),
+            y_max_m=_optional_float(
+                sar3d_object,
+                "y_max_m",
+                gui.processing.sar3d.y_max_m,
+                "gui.processing.sar3d",
+            ),
+            z_min_m=_optional_float(
+                sar3d_object,
+                "z_min_m",
+                gui.processing.sar3d.z_min_m,
+                "gui.processing.sar3d",
+            ),
+            z_max_m=_optional_float(
+                sar3d_object,
+                "z_max_m",
+                gui.processing.sar3d.z_max_m,
+                "gui.processing.sar3d",
+            ),
+            voxel_x_m=_optional_float(
+                sar3d_object,
+                "voxel_x_m",
+                gui.processing.sar3d.voxel_x_m,
+                "gui.processing.sar3d",
+            ),
+            voxel_y_m=_optional_float(
+                sar3d_object,
+                "voxel_y_m",
+                gui.processing.sar3d.voxel_y_m,
+                "gui.processing.sar3d",
+            ),
+            voxel_z_m=_optional_float(
+                sar3d_object,
+                "voxel_z_m",
+                gui.processing.sar3d.voxel_z_m,
+                "gui.processing.sar3d",
+            ),
+            norm_edge_margin_m=_optional_float(
+                sar3d_object,
+                "norm_edge_margin_m",
+                gui.processing.sar3d.norm_edge_margin_m,
+                "gui.processing.sar3d",
+            ),
+            smooth_sigma=_optional_float(
+                sar3d_object,
+                "smooth_sigma",
+                gui.processing.sar3d.smooth_sigma,
+                "gui.processing.sar3d",
+            ),
+            comp_range_power=_optional_float(
+                sar3d_object,
+                "comp_range_power",
+                gui.processing.sar3d.comp_range_power,
+                "gui.processing.sar3d",
+            ),
+            tx_per_segment=_optional_int(
+                sar3d_object,
+                "tx_per_segment",
+                gui.processing.sar3d.tx_per_segment,
                 "gui.processing.sar3d",
             ),
-            grid_nx=_optional_int(sar3d_object, "grid_nx", gui.processing.sar3d.grid_nx, "gui.processing.sar3d"),
-            grid_ny=_optional_int(sar3d_object, "grid_ny", gui.processing.sar3d.grid_ny, "gui.processing.sar3d"),
-            grid_nz=_optional_int(sar3d_object, "grid_nz", gui.processing.sar3d.grid_nz, "gui.processing.sar3d"),
             use_compensation=_optional_bool(
                 sar3d_object,
                 "use_compensation",
@@ -628,8 +691,18 @@ def gui_profile_from_dict(payload: dict[str, Any]) -> GuiProfileModel:
         raise ValueError("gui.processing.sar3d.frame_radius must be >= 0")
     if not 0.0 <= gui.processing.sar3d.edge_frame_weight <= 1.0:
         raise ValueError("gui.processing.sar3d.edge_frame_weight must be within [0, 1]")
-    if min(gui.processing.sar3d.grid_nx, gui.processing.sar3d.grid_ny, gui.processing.sar3d.grid_nz) < 2:
-        raise ValueError("gui.processing.sar3d.grid_nx/ny/nz must each be >= 2")
+    if min(gui.processing.sar3d.voxel_x_m, gui.processing.sar3d.voxel_y_m, gui.processing.sar3d.voxel_z_m) <= 0.0:
+        raise ValueError("gui.processing.sar3d.voxel_x/y/z_m must each be > 0")
+    if gui.processing.sar3d.x_max_m <= gui.processing.sar3d.x_min_m:
+        raise ValueError("gui.processing.sar3d.x_max_m must be > x_min_m")
+    if gui.processing.sar3d.y_max_m <= gui.processing.sar3d.y_min_m:
+        raise ValueError("gui.processing.sar3d.y_max_m must be > y_min_m")
+    if gui.processing.sar3d.z_max_m <= gui.processing.sar3d.z_min_m:
+        raise ValueError("gui.processing.sar3d.z_max_m must be > z_min_m")
+    if gui.processing.sar3d.ground_depth_m < 0.0:
+        raise ValueError("gui.processing.sar3d.ground_depth_m must be >= 0")
+    if gui.processing.sar3d.tx_per_segment < 1:
+        raise ValueError("gui.processing.sar3d.tx_per_segment must be >= 1")
 
     data_actions_object = _as_dict(gui_object.get("data_actions"), "gui.data_actions")
     gui.data_actions = GuiDataActionsStateModel(
@@ -795,14 +868,22 @@ def gui_profile_to_dict(model: GuiProfileModel) -> dict[str, Any]:
                 "apply_freq_phase_correction": gui.processing.sar3d.apply_freq_phase_correction,
                 "frame_radius": gui.processing.sar3d.frame_radius,
                 "edge_frame_weight": gui.processing.sar3d.edge_frame_weight,
+                "pitch_override_enabled": gui.processing.sar3d.pitch_override_enabled,
                 "pitch_deg": gui.processing.sar3d.pitch_deg,
-                "center_depth_m": gui.processing.sar3d.center_depth_m,
-                "half_width_x_m": gui.processing.sar3d.half_width_x_m,
-                "half_width_y_m": gui.processing.sar3d.half_width_y_m,
-                "half_width_z_m": gui.processing.sar3d.half_width_z_m,
-                "grid_nx": gui.processing.sar3d.grid_nx,
-                "grid_ny": gui.processing.sar3d.grid_ny,
-                "grid_nz": gui.processing.sar3d.grid_nz,
+                "ground_depth_m": gui.processing.sar3d.ground_depth_m,
+                "x_min_m": gui.processing.sar3d.x_min_m,
+                "x_max_m": gui.processing.sar3d.x_max_m,
+                "y_min_m": gui.processing.sar3d.y_min_m,
+                "y_max_m": gui.processing.sar3d.y_max_m,
+                "z_min_m": gui.processing.sar3d.z_min_m,
+                "z_max_m": gui.processing.sar3d.z_max_m,
+                "voxel_x_m": gui.processing.sar3d.voxel_x_m,
+                "voxel_y_m": gui.processing.sar3d.voxel_y_m,
+                "voxel_z_m": gui.processing.sar3d.voxel_z_m,
+                "norm_edge_margin_m": gui.processing.sar3d.norm_edge_margin_m,
+                "smooth_sigma": gui.processing.sar3d.smooth_sigma,
+                "comp_range_power": gui.processing.sar3d.comp_range_power,
+                "tx_per_segment": gui.processing.sar3d.tx_per_segment,
                 "use_compensation": gui.processing.sar3d.use_compensation,
                 "normalize_pairs": gui.processing.sar3d.normalize_pairs,
                 "bottom_view_deg": gui.processing.sar3d.bottom_view_deg,

python_app/models/gui_profile_schema.py

@@ -128,31 +128,39 @@ class GuiSar3dStateModel:
     """UI-only defaults for the 3D SAR (sar3d) live settings.
 
     The band, depth gate, antenna selection and motion knobs mirror the shared
-    gpr_* live-config fields; the sar3d_* knobs mirror the VNA_AlGo gpr_core
-    defaults used by the C++ sar3d processor.
+    gpr_* live-config fields; the sar3d_* knobs mirror the reference notebook
+    (Multi_antenna_3D_clean_v1.ipynb) defaults used by the C++ sar3d processor.
     """
 
     input_positions: str = ""
     output_positions: str = ""
-    min_depth_m: float = 2.0
-    max_depth_m: float = 14.0
-    start_freq_mhz: float = 3000.0
+    min_depth_m: float = 2.5
+    max_depth_m: float = 8.0
+    start_freq_mhz: float = 2000.0
     stop_freq_mhz: float = 6000.0
     speed_m_s: float = 0.0
     ignore_socket_speed_enabled: bool = False
-    look_angle_deg: float = 0.0
+    look_angle_deg: float = 17.5
     direction_sign: float = 1.0
     apply_freq_phase_correction: bool = True
-    frame_radius: int = 6
-    edge_frame_weight: float = 0.75
-    pitch_deg: float = 6.0
-    center_depth_m: float = 5.6
-    half_width_x_m: float = 1.8
-    half_width_y_m: float = 1.0
-    half_width_z_m: float = 1.8
-    grid_nx: int = 150
-    grid_ny: int = 120
-    grid_nz: int = 160
+    frame_radius: int = 0
+    edge_frame_weight: float = 0.8
+    pitch_override_enabled: bool = False
+    pitch_deg: float = 17.5
+    ground_depth_m: float = 1.25
+    x_min_m: float = -2.0
+    x_max_m: float = 2.0
+    y_min_m: float = -0.5
+    y_max_m: float = 2.0
+    z_min_m: float = 0.0
+    z_max_m: float = 8.0
+    voxel_x_m: float = 0.025
+    voxel_y_m: float = 0.02
+    voxel_z_m: float = 0.02
+    norm_edge_margin_m: float = 0.3
+    smooth_sigma: float = 1.5
+    comp_range_power: float = 0.3
+    tx_per_segment: int = 2
     use_compensation: bool = True
     normalize_pairs: bool = True
     # Bottom-view projection angle around the vertical Y axis, degrees:

python_app/orchestration/live_processing_config.py

@@ -72,19 +72,35 @@ class ProcessingLiveConfig:
     gpr_visible_x_max_m: float = 2.0
     gpr_visible_z_min_m: float = 0.0
     gpr_visible_z_max_m: float = 14.0
-    # 3D SAR (sar3d processor) parameters; defaults mirror VNA_AlGo gpr_core.
-    # The frequency band, depth gate, antenna selection and motion parameters
-    # are shared with the gpr_* fields above.
-    sar3d_frame_radius: int = 6
-    sar3d_edge_frame_weight: float = 0.75
-    sar3d_pitch_deg: float = 6.0
-    sar3d_center_depth_m: float = 5.6
-    sar3d_half_width_x_m: float = 1.8
-    sar3d_half_width_y_m: float = 1.0
-    sar3d_half_width_z_m: float = 1.8
-    sar3d_grid_nx: int = 150
-    sar3d_grid_ny: int = 120
-    sar3d_grid_nz: int = 160
+    # 3D SAR (sar3d processor) parameters; defaults mirror the reference
+    # notebook Multi_antenna_3D_clean_v1.ipynb. The frequency band, depth gate,
+    # antenna selection and motion parameters are shared with the gpr_* fields.
+    sar3d_frame_radius: int = 0
+    sar3d_edge_frame_weight: float = 0.8
+    # Frame pitch override; disabled -> the pitch follows gpr_look_angle_deg.
+    sar3d_pitch_override_enabled: bool = False
+    sar3d_pitch_deg: float = 17.5
+    # Height of the lower antenna row above the ground (scene Y=0 = ground).
+    sar3d_ground_depth_m: float = 1.25
+    # Scene grid bounds (metres): X across, Y height above the ground, Z along
+    # the ground away from the radar. Voxels outside the radar depth gate are
+    # zeroed by the mask regardless of the Z window.
+    sar3d_x_min_m: float = -2.0
+    sar3d_x_max_m: float = 2.0
+    sar3d_y_min_m: float = -0.5
+    sar3d_y_max_m: float = 2.0
+    sar3d_z_min_m: float = 0.0
+    sar3d_z_max_m: float = 8.0
+    sar3d_voxel_x_m: float = 0.025
+    sar3d_voxel_y_m: float = 0.02
+    sar3d_voxel_z_m: float = 0.02
+    # Volume finalization: Y-edge margin excluded from normalization, Gaussian
+    # smoothing sigma over (Z, X) planes.
+    sar3d_norm_edge_margin_m: float = 0.3
+    sar3d_smooth_sigma: float = 1.5
+    sar3d_comp_range_power: float = 0.3
+    # Fallback frame segmentation when per-trace capture stamps are absent.
+    sar3d_tx_per_segment: int = 2
     sar3d_use_compensation: bool = True
     sar3d_normalize_pairs: bool = True
     # Bottom-view projection angle around the vertical Y axis, degrees:
@@ -156,14 +172,22 @@ class ProcessingLiveConfig:
             "gpr_visible_z_max_m": float(self.gpr_visible_z_max_m),
             "sar3d_frame_radius": int(self.sar3d_frame_radius),
             "sar3d_edge_frame_weight": float(self.sar3d_edge_frame_weight),
+            "sar3d_pitch_override_enabled": bool(self.sar3d_pitch_override_enabled),
             "sar3d_pitch_deg": float(self.sar3d_pitch_deg),
-            "sar3d_center_depth_m": float(self.sar3d_center_depth_m),
-            "sar3d_half_width_x_m": float(self.sar3d_half_width_x_m),
-            "sar3d_half_width_y_m": float(self.sar3d_half_width_y_m),
-            "sar3d_half_width_z_m": float(self.sar3d_half_width_z_m),
-            "sar3d_grid_nx": int(self.sar3d_grid_nx),
-            "sar3d_grid_ny": int(self.sar3d_grid_ny),
-            "sar3d_grid_nz": int(self.sar3d_grid_nz),
+            "sar3d_ground_depth_m": float(self.sar3d_ground_depth_m),
+            "sar3d_x_min_m": float(self.sar3d_x_min_m),
+            "sar3d_x_max_m": float(self.sar3d_x_max_m),
+            "sar3d_y_min_m": float(self.sar3d_y_min_m),
+            "sar3d_y_max_m": float(self.sar3d_y_max_m),
+            "sar3d_z_min_m": float(self.sar3d_z_min_m),
+            "sar3d_z_max_m": float(self.sar3d_z_max_m),
+            "sar3d_voxel_x_m": float(self.sar3d_voxel_x_m),
+            "sar3d_voxel_y_m": float(self.sar3d_voxel_y_m),
+            "sar3d_voxel_z_m": float(self.sar3d_voxel_z_m),
+            "sar3d_norm_edge_margin_m": float(self.sar3d_norm_edge_margin_m),
+            "sar3d_smooth_sigma": float(self.sar3d_smooth_sigma),
+            "sar3d_comp_range_power": float(self.sar3d_comp_range_power),
+            "sar3d_tx_per_segment": int(self.sar3d_tx_per_segment),
             "sar3d_use_compensation": bool(self.sar3d_use_compensation),
             "sar3d_normalize_pairs": bool(self.sar3d_normalize_pairs),
             "sar3d_bottom_view_deg": float(self.sar3d_bottom_view_deg),

python_app/scripts/replay_raw_producer.py

@@ -20,6 +20,7 @@ never see time jump backwards. ``collection_id`` increases forever.
 from __future__ import annotations
 
 import argparse
+import json
 import logging
 from pathlib import Path
 import re
@@ -75,6 +76,22 @@ def _load_dataset(dataset_dir: Path, combos: list[ComboKey], expected_points: in
     timestamps_ns: list[int] = []
     for name, timestamp_ns in named:
         folder = dataset_dir / name
+        # Recorded per-trace capture windows from the snapshot's meta.json: the
+        # sar3d segmented timing model reconstructs the acquisition segments
+        # from them. Missing/partial meta leaves the zero "unmeasured" sentinel
+        # and the consumer falls back to its uniform-segment model.
+        windows: dict[tuple[int, int], tuple[int, int]] = {}
+        meta_path = folder / "meta.json"
+        if meta_path.is_file():
+            try:
+                meta = json.loads(meta_path.read_text())
+                for entry in meta.get("traces", []):
+                    windows[(int(entry["input"]), int(entry["output"]))] = (
+                        int(entry["capture_start_ns"]),
+                        int(entry["capture_end_ns"]),
+                    )
+            except (OSError, ValueError, KeyError, TypeError):
+                windows = {}
         traces: list[TraceData] = []
         for combo in combos:
             s21_path = folder / f"i{combo.input}_o{combo.output}_s21.npy"
@@ -94,12 +111,15 @@ def _load_dataset(dataset_dir: Path, combos: list[ComboKey], expected_points: in
                     f"{s21_path}: {s21.size} points, but radar.sweep.points = "
                     f"{expected_points}. Align the profile's sweep with the dataset."
                 )
+            window = windows.get((combo.input, combo.output), (0, 0))
             traces.append(
                 TraceData(
                     combo=combo,
                     frequency_hz=freq,
                     s11=np.zeros_like(s21),
                     s21=s21,
+                    capture_start_ns=window[0],
+                    capture_end_ns=window[1],
                 )
             )
         snapshots.append(traces)
@@ -114,28 +134,38 @@ def _build_collection(
     collection_id: int,
     frame_ns: int,
 ) -> SweepCollection:
-    """Wrap one preloaded snapshot in a SweepCollection at the given frame time."""
+    """Wrap one preloaded snapshot in a SweepCollection at the given frame time.
+
+    Recorded per-trace capture windows are re-anchored by the same shift as the
+    frame time, preserving the real intra-frame segment layout. Traces without
+    recorded windows get a synthetic window spanning the median frame gap.
+    """
     span_ns = dataset.median_gap_ns
     start_ns = frame_ns - span_ns // 2
     end_ns = frame_ns + span_ns // 2
+    shift_ns = frame_ns - dataset.timestamps_ns[index]
     traces = []
     for source in dataset.snapshots[index]:
+        if source.capture_end_ns > source.capture_start_ns:
+            capture = (source.capture_start_ns + shift_ns, source.capture_end_ns + shift_ns)
+        else:
+            capture = (start_ns, end_ns)
         traces.append(
             TraceData(
                 combo=source.combo,
                 frequency_hz=source.frequency_hz,
                 s11=source.s11,
                 s21=source.s21,
-                capture_start_ns=start_ns,
-                capture_end_ns=end_ns,
+                capture_start_ns=capture[0],
+                capture_end_ns=capture[1],
             )
         )
     return SweepCollection(
         collection_id=collection_id,
         monotonic_ns=frame_ns,
         traces=traces,
-        capture_start_ns=start_ns,
-        capture_end_ns=end_ns,
+        capture_start_ns=min(t.capture_start_ns for t in traces),
+        capture_end_ns=max(t.capture_end_ns for t in traces),
     )
 
 

python_app/scripts/replay_smoke_run.py

@@ -84,14 +84,28 @@ def _sar3d_live_config(profile_path: Path) -> ProcessingLiveConfig:
     live.sar3d_edge_frame_weight = float(
         sar3d.get("edge_frame_weight", live.sar3d_edge_frame_weight)
     )
+    live.sar3d_pitch_override_enabled = bool(
+        sar3d.get("pitch_override_enabled", live.sar3d_pitch_override_enabled)
+    )
     live.sar3d_pitch_deg = float(sar3d.get("pitch_deg", live.sar3d_pitch_deg))
-    live.sar3d_center_depth_m = float(sar3d.get("center_depth_m", live.sar3d_center_depth_m))
-    live.sar3d_half_width_x_m = float(sar3d.get("half_width_x_m", live.sar3d_half_width_x_m))
-    live.sar3d_half_width_y_m = float(sar3d.get("half_width_y_m", live.sar3d_half_width_y_m))
-    live.sar3d_half_width_z_m = float(sar3d.get("half_width_z_m", live.sar3d_half_width_z_m))
-    live.sar3d_grid_nx = int(sar3d.get("grid_nx", live.sar3d_grid_nx))
-    live.sar3d_grid_ny = int(sar3d.get("grid_ny", live.sar3d_grid_ny))
-    live.sar3d_grid_nz = int(sar3d.get("grid_nz", live.sar3d_grid_nz))
+    live.sar3d_ground_depth_m = float(sar3d.get("ground_depth_m", live.sar3d_ground_depth_m))
+    live.sar3d_x_min_m = float(sar3d.get("x_min_m", live.sar3d_x_min_m))
+    live.sar3d_x_max_m = float(sar3d.get("x_max_m", live.sar3d_x_max_m))
+    live.sar3d_y_min_m = float(sar3d.get("y_min_m", live.sar3d_y_min_m))
+    live.sar3d_y_max_m = float(sar3d.get("y_max_m", live.sar3d_y_max_m))
+    live.sar3d_z_min_m = float(sar3d.get("z_min_m", live.sar3d_z_min_m))
+    live.sar3d_z_max_m = float(sar3d.get("z_max_m", live.sar3d_z_max_m))
+    live.sar3d_voxel_x_m = float(sar3d.get("voxel_x_m", live.sar3d_voxel_x_m))
+    live.sar3d_voxel_y_m = float(sar3d.get("voxel_y_m", live.sar3d_voxel_y_m))
+    live.sar3d_voxel_z_m = float(sar3d.get("voxel_z_m", live.sar3d_voxel_z_m))
+    live.sar3d_norm_edge_margin_m = float(
+        sar3d.get("norm_edge_margin_m", live.sar3d_norm_edge_margin_m)
+    )
+    live.sar3d_smooth_sigma = float(sar3d.get("smooth_sigma", live.sar3d_smooth_sigma))
+    live.sar3d_comp_range_power = float(
+        sar3d.get("comp_range_power", live.sar3d_comp_range_power)
+    )
+    live.sar3d_tx_per_segment = int(sar3d.get("tx_per_segment", live.sar3d_tx_per_segment))
     live.sar3d_use_compensation = bool(sar3d.get("use_compensation", live.sar3d_use_compensation))
     live.sar3d_normalize_pairs = bool(sar3d.get("normalize_pairs", live.sar3d_normalize_pairs))
     live.sar3d_bottom_view_deg = float(sar3d.get("bottom_view_deg", live.sar3d_bottom_view_deg))

python_app/scripts/sar3d_reference_check.py

@@ -0,0 +1,389 @@
+"""Reference implementation of the sar3d volume from Multi_antenna_3D_clean_v1.ipynb.
+
+Executable specification for the C++ port: recomputes a single-frame
+(aperture radius 0) 3D SAR volume from a preprocessed snapshot exactly the
+way the notebook does, then compares it against a reference *_volume.npz
+exported by the notebook.
+
+Usage:
+    python sar3d_reference_check.py <dataset_root> <reference_volume.npz>
+
+<dataset_root> is an extracted replay dataset (contains config_profile.json
+and preprocessed/<snapshot>/i*_o*_{s21,freq}.npy + meta.json). The snapshot
+name and all physics parameters are taken from the npz meta.
+
+Exits non-zero when the recomputed volume differs from the reference by more
+than the tolerance. The only expected difference is the Gaussian smoothing
+(manual separable filter vs scipy) at the 1e-6 level.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+
+TOLERANCE = 2e-6
+
+C_LIGHT = 3e8
+
+# Notebook constants not recorded in the npz meta (cell 0 / cell 4).
+BP_OVERSAMPLE = 8
+BP_WINDOW = True
+PAIR_NORM_PERCENTILE = 50.0
+PAIR_NORM_EPS = 1e-15
+COMP_RANGE_POWER = 0.3
+COMP_RANGE_WEIGHT_MAX = 5.0
+COMP_REF_DEPTH = 5.0
+BP_SMOOTH_SIGMA = 1.5
+SAR_Y_RANGE_M = (-0.5, 2.0)
+SAR_NORM_EDGE_MARGIN_M = 0.3
+X_MARGIN_M = 2.0
+MOTION_PHASE_SIGN = -1.0
+DIRECTION_SIGN = +1.0
+
+
+def gaussian_kernel1d(sigma: float, radius: int) -> np.ndarray:
+    x = np.arange(-radius, radius + 1, dtype=float)
+    phi = np.exp(-0.5 / (sigma * sigma) * x * x)
+    return phi / phi.sum()
+
+
+def correlate1d_reflect(arr: np.ndarray, weights: np.ndarray, axis: int) -> np.ndarray:
+    r = (len(weights) - 1) // 2
+    a = np.moveaxis(arr, axis, -1)
+    pad = np.concatenate([a[..., r - 1::-1], a, a[..., :-r - 1:-1]], axis=-1)
+    out = np.zeros_like(a)
+    for k, w in enumerate(weights):
+        out += w * pad[..., k:k + a.shape[-1]]
+    return np.moveaxis(out, -1, axis)
+
+
+def gaussian_filter(arr: np.ndarray, sigma: tuple[float, ...], truncate: float = 4.0) -> np.ndarray:
+    """scipy.ndimage.gaussian_filter equivalent (mode='reflect')."""
+    out = np.asarray(arr, dtype=float).copy()
+    for axis, s in enumerate(sigma):
+        if s <= 0:
+            continue
+        radius = int(truncate * float(s) + 0.5)
+        out = correlate1d_reflect(out, gaussian_kernel1d(float(s), radius), axis)
+    return out
+
+
+def load_geometry(profile_path: Path) -> dict:
+    profile = json.loads(profile_path.read_text())
+    gpr = profile["gpr"]
+    tx_geom = {int(e["output_pos"]): (e["x_m"], e.get("y_m", 0.0), e.get("z_m", 0.0))
+               for e in gpr["tx_geometry"]}
+    rx_geom = {int(e["input_pos"]): (e["x_m"], e.get("y_m", 0.0), e.get("z_m", 0.0))
+               for e in gpr["rx_geometry"]}
+    combos = profile["run"]["combos"]
+    pair_order = [(int(c["output"]), int(c["input"])) for c in combos]
+    tx_order_full: list[int] = []
+    for tx, _ in pair_order:
+        if tx not in tx_order_full:
+            tx_order_full.append(tx)
+    output_idx = sorted(tx_geom)
+    input_idx = sorted(rx_geom)
+    tx_xyz = np.array([tx_geom[i] for i in output_idx], dtype=float)
+    rx_xyz = np.array([rx_geom[i] for i in input_idx], dtype=float)
+    return {
+        "output_idx": output_idx, "input_idx": input_idx,
+        "tx_xyz": tx_xyz, "rx_xyz": rx_xyz,
+        "pair_order": pair_order, "tx_order_full": tx_order_full,
+    }
+
+
+def snapshot_timing(snapshot_dir: Path) -> dict | None:
+    """Per-Tx capture windows from the snapshot's meta.json trace labels."""
+    try:
+        meta = json.loads((snapshot_dir / "meta.json").read_text())
+    except OSError:
+        return None
+    traces = [t for t in meta.get("traces", [])
+              if "capture_start_ns" in t and "capture_end_ns" in t and "output" in t]
+    if not traces:
+        return None
+    first = min(int(t["capture_start_ns"]) for t in traces)
+    last = max(int(t["capture_end_ns"]) for t in traces)
+    by_tx: dict[int, tuple[int, int]] = {}
+    for t in traces:
+        by_tx.setdefault(int(t["output"]), (int(t["capture_start_ns"]), int(t["capture_end_ns"])))
+    return {
+        "acq_s": (last - first) * 1e-9,
+        "center_abs_s": 0.5 * (first + last) * 1e-9,
+        "tx_window_s": {tx: ((a - first) * 1e-9, (b - first) * 1e-9) for tx, (a, b) in by_tx.items()},
+    }
+
+
+def pair_timestamps(geometry: dict, timing: dict, speed_m_s: float, look_angle_deg: float,
+                    wave_speed: float) -> dict:
+    """Segment/slot assignment and coarse pair dtau (notebook compute_pair_timestamps_for_snapshot)."""
+    windows = timing["tx_window_s"]
+    hw_order = [tx for tx in geometry["tx_order_full"] if tx in windows]
+    hw_order += [tx for tx in sorted(windows) if tx not in hw_order]
+    groups: dict[tuple[float, float], list[int]] = {}
+    for tx in hw_order:
+        groups.setdefault(windows[tx], []).append(tx)
+    seg_of = {}
+    for seg_idx, win in enumerate(sorted(groups)):
+        for slot, tx in enumerate(groups[win]):
+            seg_of[tx] = (seg_idx, slot, len(groups[win]))
+    cos_theta = np.cos(np.radians(look_angle_deg))
+    t_ref = 0.5 * timing["acq_s"]
+    tx_map = {p: k for k, p in enumerate(geometry["output_idx"])}
+    rx_map = {p: k for k, p in enumerate(geometry["input_idx"])}
+    ts = {}
+    for tx_phys, rx_phys in geometry["pair_order"]:
+        _, slot, n_slots = seg_of[tx_phys]
+        t0, t1 = windows[tx_phys]
+        t_center = 0.5 * (t0 + t1)
+        dz_motion = DIRECTION_SIGN * speed_m_s * (t_center - t_ref) * cos_theta
+        ts[(tx_map[tx_phys], rx_map[rx_phys])] = {
+            "slot": slot, "n_slots": n_slots,
+            "segment_start_s": t0, "segment_time_s": t1 - t0,
+            "t_center_s": t_center,
+            "dtau_motion_s": 2.0 * dz_motion / wave_speed,
+        }
+    return ts
+
+
+def timing_abs_center(snapshot_dir: Path) -> float:
+    timing = snapshot_timing(snapshot_dir)
+    if timing is None:
+        raise RuntimeError(f"meta.json trace labels required in {snapshot_dir}")
+    return timing["center_abs_s"]
+
+
+def load_mimo(snapshot_dir: Path, geometry: dict) -> tuple[dict, dict]:
+    s21_data, freq_data = {}, {}
+    for f in snapshot_dir.glob("i*_o*_s21.npy"):
+        parts = f.stem.split("_")
+        rx_phys = int(parts[0][1:])
+        tx_phys = int(parts[1][1:])
+        if tx_phys not in geometry["output_idx"] or rx_phys not in geometry["input_idx"]:
+            continue
+        key = (geometry["output_idx"].index(tx_phys), geometry["input_idx"].index(rx_phys))
+        s21_data[key] = np.load(f)
+        freq_data[key] = np.asarray(np.load(f.parent / f"i{rx_phys}_o{tx_phys}_freq.npy"), dtype=float)
+    return s21_data, freq_data
+
+
+def apply_phase_correction(s21: np.ndarray, freq_full: np.ndarray, info: dict,
+                           f_start: float, f_stop: float, speed_m_s: float,
+                           look_angle_deg: float, wave_speed: float) -> np.ndarray:
+    """Coarse pair + intra-sweep motion correction as one phase ramp before the IFFT."""
+    idx = np.flatnonzero((freq_full >= f_start) & (freq_full <= f_stop))
+    n_full = len(freq_full)
+    dt_event = info["segment_time_s"] / (info["n_slots"] * n_full)
+    t_abs = info["segment_start_s"] + (info["n_slots"] * idx + info["slot"]) * dt_event
+    dt_intra = t_abs - info["t_center_s"]
+    dtau_intra = 2.0 * DIRECTION_SIGN * speed_m_s * np.cos(np.radians(look_angle_deg)) * dt_intra / wave_speed
+    total_dtau = dtau_intra + info["dtau_motion_s"]
+    phi = 2.0 * np.pi * freq_full[idx] * total_dtau
+    out = np.array(s21, dtype=np.complex128, copy=True)
+    out[idx] *= np.exp(1j * MOTION_PHASE_SIGN * phi)
+    return out
+
+
+def compute_ascan(s21: np.ndarray, freq: np.ndarray, f_start: float, f_stop: float):
+    """Oversampled complex A-scan with the notebook's frequency-shifted IFFT."""
+    mask = (freq >= f_start) & (freq <= f_stop)
+    freq_cut = freq[mask]
+    s21_cut = s21[mask]
+    df = float(np.median(np.diff(freq_cut)))
+    n = len(freq_cut)
+    k0 = int(round(freq_cut[0] / df))
+    n_fft = (1 << int(np.ceil(np.log2(2 * (k0 + n - 1))))) * BP_OVERSAMPLE
+    dt = 1.0 / (n_fft * df)
+    s = s21_cut * np.hanning(n) if BP_WINDOW else s21_cut.copy()
+    spectrum = np.zeros(n_fft, dtype=np.complex128)
+    spectrum[k0:k0 + n] = s
+    return np.arange(n_fft, dtype=float) * dt, np.fft.ifft(spectrum)
+
+
+def interp_complex(tau: np.ndarray, t_axis: np.ndarray, h_axis: np.ndarray) -> np.ndarray:
+    n = t_axis.size
+    dt = (float(t_axis[-1]) - float(t_axis[0])) / (n - 1)
+    pos = (tau - float(t_axis[0])) / dt
+    inside = (pos >= 0.0) & (pos <= n - 1)
+    k = np.clip(np.floor(pos), 0, n - 2).astype(np.intp)
+    frac = pos - k
+    out = h_axis[k] + frac * (h_axis[k + 1] - h_axis[k])
+    return np.where(inside, out, 0.0 + 0.0j)
+
+
+def axis_linspace(lo: float, hi: float, step: float) -> np.ndarray:
+    return np.linspace(lo, hi, max(int(round((hi - lo) / step)) + 1, 2))
+
+
+def compute_volume(dataset_root: Path, meta: dict) -> dict:
+    """Single-frame volume via the notebook's ANIMATION path.
+
+    The reference npz files were exported from animation frames: the base grid
+    is shifted by the radar displacement between the frame and the scene-origin
+    snapshot (meta reference_snapshot), antennas get the same offset, and the
+    depth window follows the radar (depth_shift). Physically it all cancels for
+    an aperture of radius 0, but replicating the exact float arithmetic is what
+    makes the comparison bit-tight at the depth-mask knife edge.
+    """
+    geometry = load_geometry(dataset_root / "config_profile.json")
+    snapshot_dir = dataset_root / "preprocessed" / meta["snapshot"]
+
+    eps_r = 1.0
+    wave_speed = C_LIGHT / np.sqrt(eps_r)
+    f_start = float(meta["f_start_hz"])
+    f_stop = float(meta["f_stop_hz"])
+    min_depth = float(meta["min_depth_m"])
+    max_depth = float(meta["max_depth_m"])
+    speed = float(meta["speed_m_s"])
+    look_angle = float(meta["look_angle_deg"])
+    ground_depth = float(meta["ground_depth_m"])
+    voxel = [float(a) for a in meta["voxel_m"]]
+
+    x_tx, y_tx, z_tx = geometry["tx_xyz"].T
+    x_rx, y_rx, z_rx = geometry["rx_xyz"].T
+    n_tx, n_rx = len(x_tx), len(x_rx)
+
+    timing = snapshot_timing(snapshot_dir)
+    if timing is None:
+        raise RuntimeError(f"meta.json trace labels required in {snapshot_dir}")
+    ts = pair_timestamps(geometry, timing, speed, look_angle, wave_speed)
+
+    s21_data, freq_data = load_mimo(snapshot_dir, geometry)
+    ascan_h, ascan_t = {}, {}
+    for key, s21 in s21_data.items():
+        corrected = apply_phase_correction(
+            s21, freq_data[key], ts[key], f_start, f_stop, speed, look_angle, wave_speed)
+        t_axis, h = compute_ascan(corrected, freq_data[key], f_start, f_stop)
+        z_axis = t_axis * wave_speed / 2
+        m = (z_axis >= min_depth) & (z_axis <= max_depth)
+        scale = float(np.percentile(np.abs(h)[m], PAIR_NORM_PERCENTILE))
+        if not np.isfinite(scale) or scale <= PAIR_NORM_EPS:
+            scale = 1.0
+        ascan_h[key] = h / (scale + PAIR_NORM_EPS)
+        ascan_t[key] = t_axis
+
+    pitch = np.radians(look_angle)
+    cos_p, sin_p = np.cos(pitch), np.sin(pitch)
+
+    def scene_to_radar_yz(y_scene, z_scene):
+        y0 = y_scene - ground_depth
+        return y0 * cos_p + z_scene * sin_p, -y0 * sin_p + z_scene * cos_p
+
+    x_ant = np.concatenate([x_tx, x_rx])
+    y_lo, y_hi = SAR_Y_RANGE_M
+    z_span = [(d - float(scene_to_radar_yz(y, 0.0)[1])) / cos_p
+              for y in (y_lo, y_hi) for d in (min_depth, max_depth)]
+    gx = axis_linspace(float(x_ant.min() - X_MARGIN_M), float(x_ant.max() + X_MARGIN_M), voxel[0])
+    gy = axis_linspace(y_lo, y_hi, voxel[1])
+    gz = axis_linspace(min(z_span), max(z_span), voxel[2])
+
+    # Displacement of this frame relative to the scene-origin snapshot
+    # (notebook sar_frame_offsets + anim_frame_displacement).
+    off_dy = off_dz = 0.0
+    disp_y = disp_z = 0.0
+    reference_name = meta.get("reference_snapshot", meta["snapshot"])
+    if reference_name != meta["snapshot"]:
+        ref_timing = snapshot_timing(dataset_root / "preprocessed" / reference_name)
+        if ref_timing is None:
+            raise RuntimeError(f"meta.json trace labels required in {reference_name}")
+        dt = timing_abs_center(snapshot_dir) - ref_timing["center_abs_s"]
+        horizontal = DIRECTION_SIGN * speed * dt
+        off_dy = float(horizontal * np.sin(pitch))
+        off_dz = float(horizontal * np.cos(pitch))
+        # Scene displacement = difference of two transformed points; the ground
+        # lift cancels, leaving essentially (0, ~0, horizontal).
+        disp_y = (off_dy * cos_p - off_dz * sin_p + ground_depth) - ground_depth
+        disp_z = off_dy * sin_p + off_dz * cos_p
+    gy_s = gy + disp_y
+    gz_s = gz + disp_z
+    depth_shift = off_dz
+    grid_xx, grid_zz = np.meshgrid(gx, gz_s)
+
+    ground_plane_y = (-ground_depth + COMP_REF_DEPTH * sin_p) / cos_p
+
+    def pair_geo_ref(i, j):
+        xc = (x_tx[i] + x_rx[j]) / 2.0
+        r_tx = np.sqrt((xc - x_tx[i]) ** 2 + (ground_plane_y - y_tx[i]) ** 2 + (COMP_REF_DEPTH - z_tx[i]) ** 2)
+        r_rx = np.sqrt((xc - x_rx[j]) ** 2 + (ground_plane_y - y_rx[j]) ** 2 + (COMP_REF_DEPTH - z_rx[j]) ** 2)
+        return 1.0 / (r_tx * r_rx + 1e-12) + 1e-30
+
+    geo_ref = {(i, j): pair_geo_ref(i, j) for i in range(n_tx) for j in range(n_rx)}
+    dx2_tx = [(grid_xx - x_tx[i]) ** 2 for i in range(n_tx)]
+    dx2_rx = [(grid_xx - x_rx[j]) ** 2 for j in range(n_rx)]
+
+    acc_vol = np.zeros((gy.size, gz.size, gx.size), dtype=np.complex128)
+    cnt_vol = np.zeros(acc_vol.shape, dtype=float)
+    for iy, y_plane in enumerate(gy_s):
+        y_rad, z_rad = scene_to_radar_yz(y_plane, grid_zz)
+        depth_ok = (z_rad - depth_shift >= min_depth) & (z_rad - depth_shift <= max_depth)
+        r_tx_all = [np.sqrt(dx2_tx[i] + (y_rad - (y_tx[i] + off_dy)) ** 2 + (z_rad - (z_tx[i] + off_dz)) ** 2)
+                    for i in range(n_tx)]
+        r_rx_all = [np.sqrt(dx2_rx[j] + (y_rad - (y_rx[j] + off_dy)) ** 2 + (z_rad - (z_rx[j] + off_dz)) ** 2)
+                    for j in range(n_rx)]
+        acc = np.zeros_like(grid_xx, dtype=np.complex128)
+        cnt = np.zeros_like(grid_xx, dtype=float)
+        for i in range(n_tx):
+            for j in range(n_rx):
+                key = (i, j)
+                if key not in ascan_h:
+                    continue
+                r_tx, r_rx = r_tx_all[i], r_rx_all[j]
+                tau = (r_tx + r_rx) / wave_speed
+                t_axis = ascan_t[key]
+                valid = depth_ok & (tau >= t_axis[0]) & (tau <= t_axis[-1])
+                h = interp_complex(tau, t_axis, ascan_h[key])
+                ratio = (1.0 / (r_tx * r_rx + 1e-12) + 1e-30) / geo_ref[key]
+                weight = np.clip(1.0 / (ratio ** COMP_RANGE_POWER + 1e-12), 0.0, COMP_RANGE_WEIGHT_MAX)
+                acc += np.where(valid, h * weight, 0.0 + 0.0j)
+                cnt += valid
+        acc_vol[iy] = acc
+        cnt_vol[iy] = cnt
+
+    volume = np.abs(acc_vol / (cnt_vol + 1e-12))
+    z_rad_planes = np.asarray([scene_to_radar_yz(yp, grid_zz)[1] - depth_shift for yp in gy_s])
+    volume = np.where((z_rad_planes >= min_depth) & (z_rad_planes <= max_depth), volume, 0.0)
+    ref_planes = (gy_s >= gy_s[0] + SAR_NORM_EDGE_MARGIN_M) & (gy_s <= gy_s[-1] - SAR_NORM_EDGE_MARGIN_M)
+    level_raw = volume[ref_planes].max() + 1e-12
+    volume = gaussian_filter(volume / level_raw, sigma=(0.0, BP_SMOOTH_SIGMA, BP_SMOOTH_SIGMA))
+    level_smooth = volume[ref_planes].max() + 1e-12
+    volume = np.clip(volume / level_smooth, 0.0, None).astype(np.float32)
+    return {"volume": volume, "level": float(level_raw * level_smooth), "x": gx, "y": gy, "z": gz}
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    parser.add_argument("dataset_root", type=Path,
+                        help="extracted dataset dir (config_profile.json + preprocessed/)")
+    parser.add_argument("reference_npz", type=Path, help="reference *_volume.npz from the notebook")
+    args = parser.parse_args()
+
+    ref = np.load(args.reference_npz)
+    meta = json.loads(str(ref["meta"]))
+    print(f"snapshot: {meta['snapshot']}")
+    result = compute_volume(args.dataset_root, meta)
+
+    volume, ref_volume = result["volume"], ref["volume"]
+    axes_ok = all(np.allclose(result[k], ref[f"{k}_m"]) for k in ("x", "y", "z"))
+    level_rel = abs(result["level"] - meta["level"]) / abs(meta["level"])
+    print(f"axes match: {axes_ok}; level rel diff: {level_rel:.3e} "
+          f"(mine {result['level']:.9f}, ref {meta['level']:.9f})")
+    if volume.shape != ref_volume.shape:
+        print(f"FAIL: shape mismatch {volume.shape} vs {ref_volume.shape}")
+        return 1
+    diff = np.abs(volume.astype(np.float64) - ref_volume.astype(np.float64))
+    argmax = np.unravel_index(int(np.argmax(volume)), volume.shape)
+    ref_argmax = np.unravel_index(int(np.argmax(ref_volume)), ref_volume.shape)
+    print(f"max |diff| = {diff.max():.3e}, mean = {diff.mean():.3e}")
+    print(f"argmax: mine {argmax}, ref {ref_argmax}")
+    ok = axes_ok and diff.max() <= TOLERANCE and argmax == ref_argmax and level_rel <= TOLERANCE
+    print("OK" if ok else "FAIL")
+    return 0 if ok else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

python_app/scripts/sar3d_volume_compare.py

@@ -0,0 +1,89 @@
+"""Compare a sar3d volume dumped by the C++ processor with a notebook npz export.
+
+Usage:
+    python sar3d_volume_compare.py <dump_base> <reference_volume.npz>
+
+<dump_base> is the path prefix written by `sar3d_parity --dump-dir` (or by the
+processor with `sar3d_volume_dump_dir` set): <dump_base>.f32 holds the raw
+float32 volume [iy][iz][ix], <dump_base>.json its axes and level.
+
+The C++ path accumulates the backprojection in float32 (CUDA kernel), so the
+comparison is tolerance-based rather than bitwise; geometry is computed in
+double, keeping the expected difference at the 1e-5 level. Voxels within the
+smoothing radius of the depth-mask boundary are reported separately: there the
+float32 knife-edge can flip whole boundary voxels (the same effect exists
+between two float64 runs with shifted grids).
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+
+MAX_DIFF_TOLERANCE = 1e-4
+LEVEL_REL_TOLERANCE = 1e-4
+# Gaussian smoothing radius (sigma 1.5 -> radius 6) around the depth-mask edge.
+EDGE_EXCLUSION_VOXELS = 7
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    parser.add_argument("dump_base", type=Path, help="path prefix of the .f32/.json dump")
+    parser.add_argument("reference_npz", type=Path, help="reference *_volume.npz from the notebook")
+    args = parser.parse_args()
+
+    header = json.loads(Path(str(args.dump_base) + ".json").read_text())
+    shape = (header["ny"], header["nz"], header["nx"])
+    volume = np.fromfile(str(args.dump_base) + ".f32", dtype=np.float32).reshape(shape)
+
+    ref = np.load(args.reference_npz)
+    ref_volume = ref["volume"]
+    meta = json.loads(str(ref["meta"]))
+    print(f"snapshot: {meta['snapshot']}")
+    if volume.shape != ref_volume.shape:
+        print(f"FAIL: shape mismatch {volume.shape} vs {ref_volume.shape}")
+        return 1
+
+    axes_ok = all(
+        np.allclose(np.asarray(header[key]), ref[key])
+        for key in ("x_m", "y_m", "z_m")
+    )
+    level_rel = abs(header["level"] - meta["level"]) / abs(meta["level"])
+    print(f"axes match: {axes_ok}; level rel diff: {level_rel:.3e} "
+          f"(dump {header['level']:.9f}, ref {meta['level']:.9f})")
+
+    diff = np.abs(volume.astype(np.float64) - ref_volume.astype(np.float64))
+
+    # Depth-mask boundary band: voxels whose (iy, iz) sit within the smoothing
+    # radius of a masked/unmasked transition. Reconstructed from the reference
+    # parameters exactly like the processor computes the mask.
+    pitch = np.radians(float(meta.get("ground_pitch_deg") or meta["look_angle_deg"]))
+    y0 = np.asarray(ref["y_m"]) - float(meta["ground_depth_m"])
+    z_rad = -y0[:, None] * np.sin(pitch) + np.asarray(ref["z_m"])[None, :] * np.cos(pitch)
+    mask = (z_rad >= float(meta["min_depth_m"])) & (z_rad <= float(meta["max_depth_m"]))
+    edge = np.zeros_like(mask)
+    r = EDGE_EXCLUSION_VOXELS
+    for shift in range(-r, r + 1):
+        edge |= np.roll(mask, shift, axis=1) != mask
+    interior = ~edge
+
+    argmax = np.unravel_index(int(np.argmax(volume)), volume.shape)
+    ref_argmax = np.unravel_index(int(np.argmax(ref_volume)), ref_volume.shape)
+    interior_diff = diff[interior[:, :, None] & np.ones(volume.shape, dtype=bool)]
+    print(f"max |diff| (all voxels)          = {diff.max():.3e}")
+    print(f"max |diff| (off depth-mask edge) = {interior_diff.max():.3e}, "
+          f"mean = {interior_diff.mean():.3e}")
+    print(f"argmax: dump {argmax}, ref {ref_argmax}")
+
+    ok = (axes_ok and level_rel <= LEVEL_REL_TOLERANCE and argmax == ref_argmax
+          and interior_diff.max() <= MAX_DIFF_TOLERANCE)
+    print("OK" if ok else "FAIL")
+    return 0 if ok else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())