Add CSG demand frontier counters

This commit is contained in:
Meta Vens 2026-05-07 08:20:25 +01:00
commit 6bb8726a37
9 changed files with 301 additions and 25 deletions

View file

@ -80,11 +80,11 @@ Repo-local Geometry Script clone notes live in
`.\tools\setup_geometry_script_clone.ps1` if `external/geometry-script` is
missing.
The current next action is to replace the ordered `vg_csg` dense rotated
subtraction path with the category-router kernel, then prototype a
demand-shaped affected-pair frontier before adding any generic BVH/grid spatial
index. Keep cached-vs-dirty output parity tests green and compare against the
C++ timing oracle.
The current next action is to use the new demand-frontier counters to replace
the ordered `vg_csg` dense rotated subtraction path with a category-router
kernel. Keep cached-vs-dirty output parity tests green, compare against the C++
timing oracle, and only add BVH/grid storage after frontier batching proves it
needs one.
Spatial-query correction from EpiphanyAquarium memory: GigaVoxels, froxels, and
Dreams all point away from static "better tree" thinking. Let the consumer
@ -93,6 +93,11 @@ tile/view, or visible surface set. BVHs, grids, sweep-and-prune, and pair caches
are candidate layouts only after that frontier exists and timing evidence says
they help.
First implementation slice: `vg_csg` exports `DemandFrontier`/`DemandPair`, and
`BuildReport` plus the JSONL perf fixture now expose `candidate_pairs` and
`rejected_pairs`. These counters are currently observational and follow existing
bounds gates; they are not yet a replacement router.
## Important Invariants
- Target Blender 5.0+ unless the task says otherwise.

View file

@ -43,6 +43,19 @@ not carry live next-action clutter.
## Near-Term Sequence
0. Make the Rust CSG machine meaner and leaner.
- Replace the dense rotated subtraction carver with a category-router path
that preserves source polygons, splits only crossings, routes categories,
and keeps cached-vs-dirty output parity green before performance claims.
- Add a demand-frontier prototype before any generic BVH/grid index. The
frontier should emit affected brush pairs from dirty brushes, branch
bounds, and requested output scope; indexes are candidate storage layouts,
not the idea.
- Add counters and fixtures for global rebuild, affected-brush rebuild, and
request/tile-scoped rebuild so timings explain what work was refused.
- Move hot polygon/category work toward reusable scratch arenas and compact
arrays only after the router proves exact output parity.
1. Promote the Bloom coordinate/noise lessons into reusable builder seams.
- Build on the new `vibegeometry/` helper library rather than adding more
one-off example functions.