Expose dirty CSG perf mode

This commit is contained in:
Meta Vens 2026-05-07 04:31:59 +01:00
commit fc93094323
8 changed files with 145 additions and 40 deletions

View file

@ -60,8 +60,9 @@ crossing polygon classification into inside/outside/aligned/reverse-aligned
buckets, polygon category/visibility/reversal metadata, primitive/solid/tree
bounds, additive `CylinderZ`, `DomeCapZ`, and `FloretArm` primitives,
axis-aligned box fast path, compiled brush geometry cache, owned convex
splitting, stable-generation output cache, release perf fixtures, Bevy-math
mesh buffers, tests, clippy, and examples.
splitting, convex containment shortcuts, stable-generation output cache, stable
and dirty release perf fixture modes, Bevy-math mesh buffers, tests, clippy,
and examples.
Research note: `docs/research/realtime-csg-bevy-assembler.md`.
Where `vg_csg` overlaps public RealtimeCSG/demo behavior, add exact observable
parity fixtures instead of approximate plausibility tests.
@ -79,10 +80,9 @@ 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 add a dirty-rebuild benchmark mode, then use the
RealtimeCSG C++ timing oracle to guide replacement of the ordered assembler
backend with the category-router kernel where cached stable output cannot hide
fragment-carving costs.
The current next action is to replace the ordered `vg_csg` dense rotated
subtraction path with the category-router kernel while keeping cached-vs-dirty
output parity tests green and comparing against the C++ timing oracle.
## Important Invariants