[CanReader]

Performance pass based on an external audit. Every claim was checked against the code first; one commit per fix.

Changes

• Render queue: commands come from per-frame arenas instead of a heap allocation plus a RefPtr control block each; no list copies per draw.
• Vulkan: a material rebound in the same frame reuses its PBR descriptor set; SSAO linearizes sample depth instead of two mat4 transforms per tap.
• Culling: SSE2 occluder raster and rect depth scan (bit-identical depth, identical visibility; scalar fallback kept).
• OpenGL: vertex layout applied once per layout change (separate attribute formats), shadow viewport from tracked state, memoized light-space caster transforms, SSAO tap depth linearized.
• D3D11/D3D12: cached rasterizer/depth-stencil states, depth-stencil state leak fixed, static_cast buffer binds, no busy spin on the stats query, memoized shadow caster transforms.
• Scene: transform world matrix cached behind a dirty flag; culled meshes submit no transform/material state; plane/cylinder factories add the transform first; Arvo cull bounds; cached component lookups.
• Physics: broadphase queried only from movable bodies, velocity-enlarged proxies, mesh narrowphase tests in place with a bounds early-out.
• Assets/animation: models parsed once; keyframes found by binary search.
• Containers/math: HashTable hashes keys by value (plus probe, duplicate-insert and double-free fixes), O(1) Queue pop, exact-size List copies and std::sort, debug-only Matrix bounds check.
• Build: Release LTO and -fno-math-errno on the Engine target.
• Bench: CPU_frame_ms benchmark column (frame CPU time excluding present).

Results (SleakCraft, Vulkan, 1600x900, render distance 12, base and new runs interleaved, medians)

	before	after
FPS (default config)	~212	~220
CPU frame	2.05 ms	1.62 ms
Occlusion raster	1.17 ms	0.70 ms
GPU frame	2.07 ms	2.09 ms
FPS with SSAO/SSR/bloom	~150-160	~184
SSAO pass	0.42 ms	0.37 ms

OpenGL: CPU frame -21%, FPS unchanged (present-bound at ~345).

Verification

• Vulkan capture of the default scene is byte-identical to staging. With SSAO on, the only difference comes from the SSAO shader (swapping the old SPIR-V back in reproduces staging exactly).
• OpenGL screenshots match staging.
• D3D11/D3D12 only syntax-checked (zig, MinGW headers, Debug and Release); not run on Windows.
• Physics, animation, model loading and the component changes were checked with standalone bit-identity harnesses but are barely exercised by SleakCraft; they should be tried in SleakSims.

Not done

• Physics Step stays in Update (moving it to FixedUpdate without render interpolation would judder the player camera above 60 FPS).
• GBuffer/bloom descriptor caching skipped (microseconds per frame, device-lost risk if invalidation is missed).
• Skinned meshes in the GBuffer pass, TRS animation blending, intrusive RefPtr, vertex format split, -march=native and /fp:fast.