bevy_pbr/src/material_bind_groups.rs chooses a fitting slab requiring the fewest new resource slots instead of first-fit allocation. It prefers already resident textures, then fuller slabs, while keeping existing allocation/refcount/free logic and capacity limits. This adds work during material allocation, not each frame. The game's explicit Vulkan shader probe tests reuse and retirement using real GPU texture identities. See docs/cpu-followup-optimizations.md for scope and limits.

Conservative occlusion depth pyramid

bevy_core_pipeline is vendored from crates.io 0.18.1 with its original licenses. Its src/experimental/mip_generation/{mod.rs,downsample_depth.wgsl} backports Bevy #22603, commit 43f4552308a29f4130c3ad71ea65d48f77daea6e. The upstream file paths changed between releases; the algorithm is the upstream fix. Power-of-two pyramid levels and conservative source sampling prevent false occlusion at arbitrary viewport sizes, including scaled targets and MSAA. The associated meshlet texture-dimension fix is included in bevy_pbr/src/meshlet/meshlet_cull_shared.wgsl. The game uses ordinary meshes, not meshlets.

The late mesh preprocessing pass in 0.18.1 also binds its indirect dispatch buffer as writable storage, which fails wgpu validation when occlusion is on. bevy_pbr/src/render/{gpu_preprocess.rs,mesh_preprocess.wgsl} now declares that binding read-only in the late pass. Its counters are plain u32 in that shader variant, preserving the same buffer layout; the early producer retains atomic counters and writable storage. No dispatch count or culling decision changes.

bevy_pbr is the unmodified crates.io 0.18.1 source except for the changes described below. Its original MIT and Apache licenses are included. Cargo selects it through the workspace [patch.crates-io] entry; the Cargo registry is untouched.

Cached lightmap bindings

Bevy 0.18.1 rebuilds a bind group for every allocated lightmap slab and every render phase on each frame, even when all bound GPU resources are unchanged. On San Vanelona, mesh binding preparation measured roughly 10 ms per frame.

Local changes are restricted to:

src/lightmap/mod.rs: each slab retains at most eight bindings. Lookup compares the actual wgpu buffer identity, binding offset and size, and layout identity. Texture insertion/removal clears the cache. FIFO eviction bounds retained old buffers when phases disappear or instance buffers resize.
src/render/mesh_bindings.rs: reuse or create the binding through that cache. The lightmap slab parameter is now mutable.
src/lightmap/cache_tests.rs: explicit GPU regression test.
src/render/mesh.rs: retain complete lookup tables for both alternating instance buffers per phase. Keys include buffer identity, offset, size, layout, and the lightmap binding revision. A slab creation, texture upload, or removal changes that revision. Tables are shared by Arc, so a cache hit avoids rebuilding and cloning thousands of individual entries. Keep only two tables per phase, and discard caches for phases that disappear.
src/render/mesh.rs: cache the IDs of meshes with morph targets when RenderAssets<RenderMesh> changes. Each phase still rebuilds its dynamic morph bindings from the current GPU resources, but skips scanning static city meshes.

Writing new instance data into an existing buffer does not require rebuilding a bind group. Changing the buffer, bound range, layout, or texture does. The patch does not skip extraction, uploads, animation, or visibility updates. It retains the existing single-texture and binding-array rendering paths and shaders.

Run the Vulkan test explicitly:

cargo test -p bevy_pbr --lib cache_tracks_binding_identity_and_texture_replacement --locked -- --ignored

It covers cloned versus replaced buffers, changed binding ranges/layouts, writes to an existing buffer, texture insertion/removal, bounded eviction, and actual bind-group construction/reuse, alternating-buffer table reuse, table revisions, and morph-target additions/replacements/removals. Regular game tests and a city screenshot/benchmark also exercise the patched renderer.

Upstream references reviewed September 6, 2026:

Bevy issue 23595 reports increasing prepare_mesh_bind_groups cost. That report points at mesh scanning; local measurements and source inspection identified lightmap binding recreation as the opportunity addressed here.
Bevy 0.19 rendering improvements explain further GPU batching and CPU overhead reductions. An engine migration is separate from this small patch; 0.19.1 source still rebuilds lightmap groups.

When upgrading Bevy, reassess this patch and rerun the GPU test, stationary and moving-camera benchmarks, and visual comparison. Do not assume an unchanged private renderer implementation across engine versions.