Release v0.7.3

This release is about documentation: a complete static documentation site, and
API documentation covering every symbol in the library down to the individual
CUDA kernel.

Features

• Documentation website — a static site served from docs/, with a showcase,
  installation and concepts guide, benchmarks, and a seven-tier API reference.
  Dark-first with a light toggle, client-side search, and a raymarched signed
  distance field in the hero.
• Complete API reference — 1018 symbols, 100% documented. The reference does
  not stop at the Python surface: it descends through the pybind11 bindings and
  host dispatchers into the 109 CUDA __global__ kernels, the 69 __device__
  helpers they call, the 65 inline math primitives, and the 28 __constant__
  topology tables.
• Qualitative results gallery — seven figures rendered from real library
  output on the bundled benchmark assets: the four extractors compared against
  ground truth on a shared field, a 64³→1024³ resolution ladder, all six sign
  determination modes, per-vertex curvature fields, normal-mode comparison, SDF
  slices, and measured narrow-band memory scaling.
• Documentation generator — stdlib-only, static: ast over the Python
  package and the pybind stub, Doxygen XML for the native tiers, normalised into
  one model so a CUDA kernel and a Python function render with the same grammar.
  --strict enforces documentation coverage and internal link integrity.
• Figure pipeline — a self-contained nvdiffrast renderer that regenerates the
  entire gallery from the bundled assets, so the figures cannot go stale.

Refactors & Build

• Documented the native layer throughout: extraction kernels with their parallel
  decomposition, launch bounds, warp divergence and atomicity hazards; the BVH,
  KD-tree, grid and triangle mesh internals; the geometric and Gaussian splatting
  primitives; the constant topology tables; and the inline vector, matrix and QEF
  math.
• Added pbdoc documentation to the previously undocumented Triangle
  predicates, and documented every bind_* registration and tensor-validating
  wrapper.
• Added class and constructor docstrings to the 24 bundled benchmark asset
  classes.
• Fixed SyntaxWarning: invalid escape sequence raised on import by two
  module docstrings containing LaTeX, which now use raw strings.
• Rewrote the README around the documentation site, with the tier breakdown, a
  collapsible feature list and a citation entry.
• Bumped version to 0.7.3.

Examples

No changes to the bundled examples in this release. The differentiable rendering
example under examples/differentiable_rendering/ is unchanged.

Notes

• Documentation: https://khoidoo.github.io/conquer3d/ (requires GitHub Pages to be
  enabled for the main branch /docs folder).
• Building from source on a system with GCC 15 or newer needs the host compiler
  pinned for nvcc, e.g. CC=$CONDA_PREFIX/bin/x86_64-conda-linux-gnu-g++ with
  NVCC_APPEND_FLAGS="-std=c++20"; PyTorch 2.8's headers do not compile under
  newer GCC at C++17.