Fix blocky WebGPU sky: view inputs refresh when the camera moves (t91)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs/agent-log.md

@@ -4,6 +4,45 @@ What agents have done, newest first. Add an entry for every piece of work: the d
 note or request it was for, what changed, the files touched, the tests run, and anything left
 open or not verified. Keep each entry short; the git history has the details.
 
+## 2026-10-28: t91 sky renders as mismatched blocks (Codex)
+
+- Reproduced the owner's cloud rectangles and stepped distant hills on NVIDIA Blackwell:
+  generated seed 424242 at 2048x1024, smooth look, 1320x916 viewport, 30 submitted frames
+  panning (400,20) to (565.3,32.47) while zooming 4 to 1.68. Inspected
+  `shots/t91-before.png`, `t91-after.png`, `t91-pixel.png` and the seven-pair
+  `t91-golden-review.png`. Clouds and hills are continuous after the fix; pixel sky intact.
+- Cause: `vn` in `gpu/smooth-shader.js` hashed shared lattice corners differently depending
+  on whether they came from the current cell or its offset neighbour. Isolated GPU readback
+  measured a 255/255 jump across an infinitesimal cell boundary. Rounding the offset corner
+  sums before `h21` reduces the maximum to 1/255. This implicates fused/reassociated hash
+  arithmetic in the shader port, not stale camera inputs. The previously frozen sky/joins
+  references also contained these seams, so their presence does not establish a 2d regression.
+  Floating-point fusion/reassociation is permitted by
+  [WGSL](https://www.w3.org/TR/WGSL/#floating-point-evaluation) and
+  [Vulkan](https://docs.vulkan.org/spec/latest/appendices/spirvenv.html#spirvenv-precision-operation).
+  No backend disassembly was captured; the precise optimizer transformation is inferred.
+- Kept dirty world uploads unchanged: a forced full upload of the broken frame was exactly
+  identical. Audited view-dependent inputs: main camera/view/zoom uniform, smooth time/surface
+  uniform, overrides and row/column tables upload each draw; colour dispatch covers the full
+  view. Pixel colour shares this path. Lighting refreshes its own uniform/grid on its selected
+  cadence and retains the matching origin between updates. Existing cadence check still passes.
+- Added two checks to `test/t_gpu.py`: actual smooth noise helpers rendered to an offscreen
+  texture and read across horizontal/vertical lattice boundaries (failed before, passes after);
+  12-frame pan/zoom paths compared to recreated view textures/light history plus full world
+  uploads, pixel/smooth with full/fast/off lighting (all six comparisons exactly equal).
+  Seven smooth golden crops intentionally refreshed after inspection: renderer 12/2, joins 2,
+  sky 12/2, lamp 12 and magma 12. Shared noise also shades surfaces, so those textures change
+  slightly. All pixel and remaining smooth references retained; tolerances unchanged.
+- Files: `src/render/gpu/smooth-shader.js`, `test/t_gpu.py`, `test/t_gpu_golden.py`, seven
+  `test/fixtures/golden/*-smooth-*.png`, `docs/TESTING.md`, this log and t91 tracking.
+  No simulation changes, DESIGN/other task edits or commit.
+- Verification: focused GPU suite passes (14 checks, no page/WebGPU validation errors).
+  Full `python3 test/run_all.py --browser` exited 0, ALL PASSED: all 32 headless suites;
+  gameplay 361, smooth 28, console 18, capabilities 7, GPU 14, goldens 28, light 71 and
+  Firefox 2 checks. Evidence: `/tmp/t91-all-tests.log`. Diff check and kandown doctor passed;
+  t91 moved to Review. Sandbox had no WebGPU adapter; host-GPU execution used for reproduction
+  and verification. Other GPUs and owner playtesting remain unverified.
+
 ## 2026-10-28: t88 phase 4 hardware verification and handoff (Codex)
 
 - Reviewed the existing uncommitted Phase 4 changes rather than duplicating the implementation.