perf(core): opt-in fast coincident-point merge for vtkCleanPolyData

[akaszynski]

Summary

Adds an opt-in fast path to vtkCleanPolyData — the filter behind pyvista's .clean() — for the common polys-only exact-merge case, reusing the already-vendored pvaClean.h (MIT) kernel via a new adapter fvtkFastCleanPoly.{h,cxx}.

Stacked on #83 (fast static-clean UG). Base is feat/fast-clean-port; review/merge #81 → #82 → #83 first. Net-new here is the vtkCleanPolyData path + cleanpoly_fast op (the kernel + OpenMP CMake come from #83).

Scope (tight, by design)

vtkCleanPolyData has rich semantics (4 cell arrays, degenerate→line→vertex conversion, strip→poly conversion). Rather than reimplement all of it, the fast path engages only for the dominant triangulated-surface case and falls back otherwise. It activates only when fvtk::FastModeEnabled() AND:

• PointMerging on, effective tolerance 0 (exact merge),
• input has polys and no verts/lines/strips,
• no point global-ids (different merge predicate) and no ghost points,
• float/double points, <2³¹ points,
• and no cell degenerates — vtkCleanPolyData would convert a collapsed poly to a line/vertex, which the kernel doesn't reproduce, so the adapter detects a repeated id in any output cell and bails to the standard path.

Everything else → standard byte-exact path (default, fast-off).

Faithful-to-stock details

• Polys kept 1:1 in input order → cell data passes through unchanged.
• Merged-point data copied from the kernel's canonical source; tolerance-0 merge → coincident points share exact coords, so output coordinates are bit-identical to stock.
• Output is polys-only (matching stock for polys-only non-degenerate input), so the captured array set matches.

Validation

• cp312-abi3 manylinux wheel; vtkFiltersCore.abi3.so links libgomp.
• Engagement proven: under EnableFast() the output point order is reordered vs the standard path (same point set, same counts) → not a silent fallback.
• Bit-exact gate: 296 passed — the 4 cleanpoly_fast cases match stock points-relaxed; the original byte-exact clean cases still pass.

🤖 Generated with Claude Code

[akaszynski]

Merged to main via fast-forward (89ebe2a). Live on main.