[paul-hammant]

Three rendering fixes, one commit each, each with a changes/2026-11-05-*.md entry. No motion code is touched.

• Closes Depth of field: rotate the sample spiral per pixel and take more samples #881
• Closes Draw count: TAA, SSR (and the depth resolve) disagree between OpenGL and Vulkan #882
• Closes World text: far, small labels shimmer (no mips; keep GL and Vulkan identical) #892

#881 Depth of field

dof_gather read the same 32 spiral taps at every texel, so the spiral's own pattern showed in the blur. The spiral is now turned at each texel by interleaved-gradient noise (the occlusion's dither, counted from the same corner on both back ends), 48 taps; radii unchanged, so an in-focus texel still takes nothing. New CPU check in test_gpu_dof: a hard edge under the widest disc, rows along the edge must differ and their mean must match an even disc's blur within 0.015 (new code 0.0077 off; main's gather 0.029 off and fails; 48 taps without the turn 0.026 off and fails). frames_against main: only the four DoF frames differ, the other 77 byte-identical. GL/Vulkan parity at 1x: 0 channels (as main). Cost at disc 100, zombie_city 1280x720: Vulkan 0.27 ms (main 0.12), OpenGL 0.35 ms (0.18); DoF is off by default, no budget pins it.

#882 Draw counts

The back ends counted different full-screen passes (OpenGL: SSR, TAA, HUD; Vulkan: depth resolve, SSAO). The rule vkframe already stated now applies everywhere: a screen pass is never a draw. draw_quoted_with no longer counts (six save/restore guards removed); OpenGL's SSR/TAA/HUD increments removed. docs/rendering.md gains "What a frame's draw count is". test_backend_parity turns on SSAO, TAA, SSR, DoF one at a time and together: every case must equal the no-post count of 7 on both back ends and match across them (5 checks fail on main). Budgets: zombie_street draw_calls 88 → 87 both back ends; zombie_city shadow-map 696 → 694 (measured on the M6 at 320x180 as ci_scenes does). Not done: the RTX 4070 Ti ray-traced zombie_city*.vulkan.budget.json (285) should now read 284 (283 with ray-traced occlusion off) and need re-recording on that card with ae3d_bench --record; noted in ci_scenes.sh and the changes entry.

#892 World text shimmer

Per the issue (no mips, keep GL and Vulkan identical), the fix is in the shader only: where a pixel spans more than one atlas texel, the pixel is covered by the mean of a grid of field reads over its footprint (up to 4x4), faded in between one and two texels per pixel; symmetric about the pixel centre so OpenGL's upward y maps onto itself. Neither atlas changed; #894 untouched. test_world_text slides the camera past a 5 px/em and a 9 px/em label for 24 frames: coverage within 2% of its mean and of the previous frame on both back ends, back ends agreeing. Main: 4.2%/6.7% and 1.9%/3.6% (4 checks fail); branch 1.3%/1.5% and 0.74%/0.76% at 4x and 1x; moving frames 0 channels apart on the M6, 9 on Intel. bench_world_text within noise.

Verification

• generate_shaders --check current (glslang 16.6.0); msl_check 60 files; test_quoted_hashes, test_vk_msl, test_gpu_msl, test_gpu_lettering pass; GLSL fixtures regenerated.
• Validation layer (sync validation on, layer confirmed loaded): parity, world text, view: 0 errors at 4x and 1x; validate.sh with DoF at its widest: 0 for zombie_city and street_life.
• cachyos (real Intel GPU, gcc 16.2 and clang 23): 0 warnings; parity and world text pass at 4x and 1x on Intel OpenGL and Vulkan, 0 layer errors.
• Local ./ci.sh: all pass except 4 "exported fixtures stale" checks (tests/fixtures/exported, resources/blender/{showcase,zombie_street,car}), whose only difference is the Blender version string (5.2.2 committed, 5.2.3 installed here); main fails the same way on this Mac. A re-export is a separate commit, not in this PR. 8 hardware skips (DLSS, ray queries, one leak report).

🤖 Generated with Claude Code