sar3d: port the Multi_antenna_3D_clean_v1 notebook algorithm

The old sar3d core was a port of the superseded Horns_3D notebook and
produced wrong volumes. The compute core now follows the reference
notebook exactly:

- segmented_interleaved frame timing rebuilt from per-trace capture
  stamps (uniform tx_per_segment split as the no-stamps fallback), with
  the coarse inter-segment dtau and the intra-sweep dtau applied as one
  phase ramp before the IFFT;
- ground scene coordinates with the ground_depth_m lift (scene Y=0 is
  the ground), full-region grid instead of a window around a center
  depth, geo_ref compensation anchored on the ground plane, power 0.3;
- notebook finalization in C++: |acc/cnt|, depth mask, normalization
  without the Y-edge planes, separable Gaussian over (Z, X), level;
- CUDA kernel computes geometry in double: float32 loses ~3e-4 of an
  A-scan sample in tau and drifts ~1e-4 from the float64 reference.

Live-config keys follow: the center_depth/half_width/grid_n window knobs
are replaced by ground/voxel/y-range/margin/sigma/power/tx_per_segment.

Parity against the notebook npz exports (frames 0034-0036 of the
20260922_mine_fast_true dataset): max|diff| <= 1.6e-6 off the depth-mask
knife edge, identical argmax, level rel diff <= 7e-7. Verified via
make sar3d_parity --dump-dir + scripts/sar3d_volume_compare.py; the
executable spec lives in scripts/sar3d_reference_check.py (bit-tight
against the npz, including the animation shifted-grid arithmetic).

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

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())