Phase 5.4. Two defects with one root cause: part_geoms was keyed on the
raw node id, so staleness was invisible to the type and correctness rested
on seventeen scattered `part_geoms.remove(&id)` calls at the edit sites.
All seventeen are deleted; the map is now keyed on (id, ParamHash), the
same key MeshCache uses, and an entry whose hash no longer matches its
node is simply never read.
1. subdivide_selected drew the wrong geometry. It replaced each selected
part's solid with a fresh Csg and invalidated the MESH cache, but never
removed the part's part_geoms entry -- so the stale uploaded buffer
stayed a hit and the viewport kept drawing the un-subdivided shape.
An audit of every mutation path found this was the only edit site
missing its manual eviction, which is exactly the failure mode a manual
protocol produces.
2. ParamHash covered the transform and the material, so a pure move
invalidated the mesh. It should not: build_mesh reads neither. The
cached TriMesh is in the node's LOCAL space, and all four consumers --
the draw loop, pick_part, the STL and glTF exporters -- apply the model
matrix themselves. Dragging a part therefore re-triangulated it on
every frame to produce byte-identical triangles, at up to 886 ns per
extruded part per frame, per selected part. The hash now covers the
solid parameters and nothing else.
Two existing tests asserted the old behaviour and were INVERTED, not
deleted -- they described what the code did rather than what it needed to
do:
cache_detects_transform_edits_via_param_hash
-> a_transform_edit_reuses_the_cached_local_space_mesh
mesh_cache_self_invalidates_on_parameter_and_transform_edits
-> mesh_cache_self_invalidates_on_solid_parameter_edits
New: build_mesh_output_does_not_depend_on_the_transform guards the
assumption the key now rests on -- if anyone makes build_mesh bake the
transform in, it fails and the key must grow it back. Plus
mesh_cache_hits_on_a_pure_transform_edit, mesh_cache_hits_on_a_material_edit
and an_uploaded_buffer_is_not_reused_after_the_solid_changes. Negative
test: restoring the transform to the hash makes the first two fail;
restored and green.
Deletion is still explicit (`retain` on the live id set) because it is the
one case a content hash cannot express -- there is no node left to hash.
ParamHash is pub(crate), not pub: it is how the caches agree on staleness,
not a consumer contract.
Also recorded in BENCH_BASELINE.md, not fixed here: bench_command_execute
_overhead is 227 us/cmd against a stale recorded 0.4 us. Verified
pre-existing -- 254 us on the commit before this branch, so this work
slightly improves it. CadCommandCtx::new eagerly builds a scene snapshot
that most commands never read, and every command bumps the generation, so
it is O(commands x nodes). The fix is to make that snapshot lazy; it is a
separate change and gets its own commit.
753 lib + 154 integration tests pass. Test-name list diffed, not just the
count. All 12 CI gates pass.
11ef0fbf67