Add narrow-band SDF reinitialize/retopologize ops

[swahtz]

Adds an SDF reinitialization operator and a narrow-band retopologization helper. The core is a VoxelBlockManager (VBM) eikonal solver ported from standalone code used in a NanoVDB meshing prototype.

Motivation

This is the first needed component of a port of the OpenVDB VolumeToMesh dual-contouring meshing tool to NanoVDB + VoxelBlockManager that I've been developing as a standalone example and I imagined is likely useful for other applications which use fvdb-core.

Changes

Core CUDA op — src/fvdb/detail/ops/ReinitializeSdf.{h,cu} (new)

reinitialize_sdf redistances a per-voxel signed field to |grad phi| = 1 on the same grid (topology unchanged):

• TVD-RK Godunov upwind eikonal solve, order 1 (forward Euler), 2 (Heun), or 3 (Shu-Osher), with a frozen Peng smoothed sign.
• Optional de-staircasing by mean-curvature or volume-preserving Taubin Laplacian smoothing (smooth passes), selected via SmoothingMode.
• The field is clamped to [-band*vx, band*vx] each sweep; inactive face-neighbours read the +band*vx "outside" value.
• Per-grid loop over the batch; per-grid value-indexed scratch buffers (slot 0 holds the background) are gathered from / scattered to the JaggedTensor.
• float32 / float64; CUDA only (the VBM is a CUDA structure — CPU raises a clear TORCH_CHECK).

Python composition — fvdb/functional/_sdf.py (new)

retopologize_sdf composes reinitialize_sdf with existing ops, no new C++ plumbing:

1. (optional, default pad=True) dilate by band via dilated_grid and carry the field onto the new voxels via inject, seeding them as exterior, so the output band is a full band voxels wide even when the input grid's active band was thinner;
2. reinitialize_sdf; (optional, default prune=True) keep |phi| < band*vx*0.999 via the existing pruned_grid and select the field with JaggedTensor.rmask. The mask-based prune preserves canonical voxel order, so the field stays aligned with the pruned grid without a re-inject.

Public API

• fvdb.functional.{reinitialize,retopologize}_sdf_{single,batch} (+ exports).
• Grid.reinitialize_sdf / Grid.retopologize_sdf and the GridBatch equivalents
• New SmoothingMode enum following fvdb's cross-boundary-enum convention

Dependency — src/cmake/get_nanovdb.cmake

• Bumped the NanoVDB pin fec59777 → f9754140 for the VoxelBlockManager build/decode API (buildVoxelBlockManager, decodeInverseMaps, VoxelBlockManagerHandle).

Test plan

Built and tested in the fvdb conda env (PyTorch 2.11, CUDA 13, sm_120):

• tests/unit/test_sdf.py: 10 passed — sphere-SDF preserved under redistancing (mean err ~0.09·vx), recovered from a crude sign step (~0.25·vx), band clamp, narrow-band prune, prune voxel-order alignment guard (grid.ijk[mask] == pruned.ijk), padding widens a thin band, batch-vs-single parity, and float64.
• NanoVDB bump regression (run before adding the op): full rebuild on the new pin with no API drift; tests/unit/test_inject.py + test_io.py (234 passed) and test_dual.py + test_basic_ops_single.py (94 passed).
• ReinitializeSdf.cu compiles clean under -Werror=all-warnings / -Xcompiler=-Wall,-Werror.

Notes / risks

• retopologize_sdf(pad=True) seeding. New voxels created with padding are seeded as exterior (+band*vx), which is correct when the dilation extends outward — i.e. the grid's interior (phi < 0) is already represented (occupancy/TSDF/mesh-derived fields, the common case). For a hollow thin shell that doesn't fill its interior, pass pad=False with a pre-banded grid. Documented in the docstring.
• Dtype. float32 / float64 only — float16 is intentionally rejected (the eikonal solve needs the precision).

[Copilot]

Pull request overview

This PR adds a CUDA-only narrow-band SDF reinitialization operator (Godunov upwind eikonal solve with optional Laplacian de-staircasing) plus a Python-level retopologize_sdf helper that composes padding/injection, reinitialization, and band pruning to produce a clean narrow-band SDF aligned with the output grid topology.

Changes:

• Add reinitialize_sdf CUDA op backed by NanoVDB VoxelBlockManager and expose it through the C++/pybind layer.
• Add Python functional + OO APIs for reinitialize_sdf and retopologize_sdf (single + batch), and introduce a cross-boundary SmoothingMode enum.
• Add focused unit tests for correctness, pruning alignment, padding behavior, batch parity, and float64.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
src/fvdb/detail/ops/ReinitializeSdf.h	Declares the new CUDA SDF reinitialization API + SmoothingMode.
src/fvdb/detail/ops/ReinitializeSdf.cu	Implements the VBM-based narrow-band redistancing + optional smoothing.
src/python/GridBatchOps.cpp	Binds SmoothingMode and reinitialize_sdf into _fvdb_cpp.
fvdb/functional/_sdf.py	Adds Python functional wrappers and retopologize_sdf composition logic.
fvdb/functional/__init__.py	Exports the new SDF functional APIs.
fvdb/enums.py	Adds public SmoothingMode IntEnum.
fvdb/grid.py	Adds Grid.reinitialize_sdf / Grid.retopologize_sdf.
fvdb/grid_batch.py	Adds GridBatch.reinitialize_sdf / GridBatch.retopologize_sdf.
fvdb/_fvdb_cpp.pyi	Adds type stubs for _fvdb_cpp.SmoothingMode and reinitialize_sdf.
fvdb/__init__.py	Re-exports SmoothingMode at top-level.
tests/unit/test_sdf.py	New unit tests for the SDF reinit/retopo behavior.
src/CMakeLists.txt	Registers the new CUDA source file in the build.
src/cmake/get_nanovdb.cmake	Updates NanoVDB pin for required VBM APIs.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 13 out of 13 changed files in this pull request and generated no new comments.

[harrism]

Looks good. Just had one nit.