Delaunay32 0.7.0

This release improves numeric edge-case handling, internal modularity, sparse polygon sampling, and SVG construction.

Migration

• Rebuild all consumers: Triangulator has a new binary layout. Public method
  signatures remain unchanged. Versioned CMake consumers should request 0.7.
• Seeded sampling remains reproducible within this release, but sparse
  multi-domain sample sequences differ from earlier releases.

Added

• Add delaunay_workload_benchmark for constraints, full results, SVG construction
  and serialization, and rectangular/disconnected polygon sampling.

Changed

• Sample sparse polygon unions through area-weighted domain bounds, correcting
  for overlapping proposal boxes. Dense regions retain global rejection
  sampling. Seeds remain deterministic, but sparse multi-domain sequences change.
• Store SVG path coordinates contiguously with ring offsets, removing per-triangle
  allocations while preserving serialized output.
• Hide triangulation implementation details behind a private implementation,
  with dedicated edge-arena storage/allocation and grouped constraint workspace.
  Public method signatures are unchanged; consumers must rebuild because the
  Triangulator binary layout changed.
• Share SVG geometry construction and polygon query traversal between integer
  and floating-point overloads, retaining their distinct arithmetic predicates.
• Preserve the topology kernel's specialized merge loops and direct storage
  access; moved-from triangulators can be configured for a new problem.

Fixed

• Blue-noise sampling falls back to linear neighbor queries when tiny bounds
  cannot support finite grid reciprocals, avoiding invalid integer conversions.
• Automatic quantization rejects nonzero spans that require an infinite scale
  instead of returning an invalid mapping and zero error report.
• Jittered polygon sampling computes ring areas relative to a local origin,
  preserving small domains translated to large coordinates.
• Sampling rejects overflowing axis spans for bounds and polygon regions
  before they can produce non-finite random coordinates.

Validation

• All 11 local Release tests passed with compiler warnings treated as errors.
• Installed-package consumer and 0.6/0.7 compatibility checks passed.
• Core-only shared-library build and full triangulation/workload benchmarks passed.
• All six release-commit CI jobs passed: Linux, macOS, Windows, AddressSanitizer/UndefinedBehaviorSanitizer, ThreadSanitizer, and forced 64-bit predicates.

Performance depends on the workload; sparse-domain sample sequences have changed as described above.

Delaunay32 0.6.3

Delaunay32 0.6.3 improves performance and fixes correctness and recovery issues. The public API, result format, and C++17 requirement remain unchanged.

Improvements

• Faster serial Morton sorting with fused radix histograms and fewer passes.
• Faster parallel topology construction through reduced cache-line sharing and reuse of child hull extrema.
• Spatial indexing accelerates validation of larger polygon boundaries.
• Full-result export builds adjacency in triangle order and inlines face discovery without adding a scratch buffer.

Fixes

• Polygon clipping correctly excludes regions separated by standalone constraints outside polygons or inside holes.
• Edge-arena allocation failures leave triangulators safe to reset and retry.
• Parallel allocation rejects blocks too small to hold a complete edge and uses the existing serial fallback.
• JSON and SVG output use locale-independent numeric formatting.

Validation covers Linux, macOS, Windows, address/undefined-behavior and thread sanitizers, and forced 64-bit predicates. Regression tests also cover sparse full results, reuse after collinear input, and output-allocation failure recovery.