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

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

docs/TESTING.md

@@ -79,17 +79,21 @@ checks waking within eight ticks separately from the probabilistic consumption o
   `shots/play/`. Pixel look by default; `python3 test/play.py shots/play smooth` uses the smooth
   look.
 - `test/t_smooth.py` checks WebGPU smooth shapes, seams, joins, fog and the thermal lens.
-- `test/t_gpu_golden.py` compares 28 cropped legacy reference PNGs (384x288, under
+- `test/t_gpu_golden.py` compares 28 cropped frozen reference PNGs (384x288, under
   `test/fixtures/golden/`) against WebGPU: renderer, joins, sky and four light scenes,
   both looks at zoom 12 and 2. ImageMagick decodes RGB; the per-channel tolerance is 8,
   mean RGB error must stay below 0.5/255 and fewer than 0.3% of pixels may exceed 8
-  (smooth AO/edge rounding). These are frozen pre-removal legacy images, not generated
+  (smooth AO/edge rounding). Seven smooth crops were refreshed for t91 (renderer 12/2,
-  from the new renderer. The compressed `golden-world-640x*.json.gz` fixtures retain the
+  joins 2, sky 12/2, lamp 12, magma 12): the old noise had seams in the references too.
-  Phase 2 world bytes that produced them, isolating renderer checks from worldgen RNG changes. Inspect the matching `shots/golden-*.png` after visible changes.
+  The remaining crops are frozen pre-removal legacy images. The compressed
+  `golden-world-640x*.json.gz` fixtures retain the Phase 2 world bytes that produced them,
+  isolating renderer checks from worldgen RNG changes. Inspect the matching
+  `shots/golden-*.png` after visible changes.
 - `test/t_gpu.py` checks default limits (four storage textures per stage), dirty-upload
-  readback after 240 busy ticks with scrolling and both looks, exact dirty/full image equality
+  readback after 240 busy ticks with scrolling and both looks, exact dirty/full image equality,
-  with a frozen clock, world replacement, transparent overlays, drill contours, shader compilation timing, successful device recreation and failed recovery with
+  noise continuity and pan/zoom equality against fresh view textures/full uploads with a
-  an autosave and Reload. `gpuCheckUploads()` reads actual resident GPU texture bytes and
+  frozen clock, world replacement, transparent overlays, drill contours, shader compilation
+  timing, successful device recreation and failed recovery with an autosave and Reload. `gpuCheckUploads()` reads actual resident GPU texture bytes and
   compares them to the full CPU arrays, including Float32 temperatures, machine and block data.
 - `test/t_render_dirty.js` checks invalidation of sleeping chunks, heat edges, swaps, fog and trees.
 - The old CPU light checks now run as `test/t_gpu_light.py` / `test/light-cases.js`, reading the

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.

src/render/gpu/smooth-shader.js

@@ -15,7 +15,8 @@ fn modulo(x: f32,y: f32) -> f32 { return x-y*floor(x/y); }
 fn h21(inputP: vec2f) -> f32 { var p=inputP; p = fract(p * vec2f(123.34, 456.21)); p += dot(p, p + 45.32); return fract(p.x * p.y); }
 fn vn(p: vec2f) -> f32 {
   var i: vec2f = floor(p); var f: vec2f = fract(p); f = f * f * (3.0 - 2.0 * f);
-  return mix(mix(h21(i), h21(i + vec2f(1, 0)), f.x), mix(h21(i + vec2f(0, 1)), h21(i + vec2f(1, 1)), f.x), f.y);
+  // Round corner sums before hashing so fused arithmetic cannot give shared corners different hashes.
+  return mix(mix(h21(i), h21(floor(i + vec2f(1, 0))), f.x), mix(h21(floor(i + vec2f(0, 1))), h21(floor(i + vec2f(1, 1))), f.x), f.y);
 }
 fn fbm(inputP: vec2f) -> f32 { var p=inputP; var a: f32 = 0.5; var s: f32 = 0.0; for (var k: i32 = 0; k < 4; k++) { s += a * vn(p); p = p * 2.03 + 17.1; a *= 0.5; } return s / 0.9375; }
 

test/t_gpu.py

@@ -57,6 +57,61 @@ with launch(viewport={'width':960,'height':600}) as (_,pg,url):
     }""")
     assert all(r['max']==0 for r in parity),parity
     print('PASS dirty and forced full uploads render identically in',len(parity),'look/light/lens/fog configurations')
+    # Adjacent noise cells must use the same hash for their shared lattice corners.
+    noise=pg.evaluate("""async()=>{
+      const code=GPU_SMOOTH.slice(GPU_SMOOTH.indexOf('fn h21'),GPU_SMOOTH.indexOf('fn lin'))+`
+      @vertex fn vs(@builtin(vertex_index) i:u32)->@builtin(position) vec4f {
+        let p=array<vec2f,3>(vec2f(-1,-1),vec2f(3,-1),vec2f(-1,3));return vec4f(p[i],0,1);
+      }
+      @fragment fn fs(@builtin(position) p:vec4f)->@location(0) vec4f {
+        let edge=floor(p.x/2.);let along=floor(p.y)*0.1;
+        let epsilon=select(-0.0001,0.0001,u32(p.x)%2u==1u);
+        return vec4f(vn(vec2f(edge+epsilon,along)),vn(vec2f(along,edge+epsilon)),0,1);
+      }`;
+      const module=GPU.device.createShaderModule({code});
+      const pipeline=await GPU.device.createRenderPipelineAsync({layout:'auto',
+        vertex:{module,entryPoint:'vs'},fragment:{module,entryPoint:'fs',targets:[{format:'rgba8unorm'}]}});
+      const texture=GPU.device.createTexture({size:[960,100],format:'rgba8unorm',
+        usage:GPUTextureUsage.RENDER_ATTACHMENT|GPUTextureUsage.COPY_SRC});
+      const buffer=GPU.device.createBuffer({size:960*100*4,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});
+      const encoder=GPU.device.createCommandEncoder();
+      const pass=encoder.beginRenderPass({colorAttachments:[{
+        view:texture.createView(),loadOp:'clear',storeOp:'store'}]});
+      pass.setPipeline(pipeline);pass.draw(3);pass.end();
+      encoder.copyTextureToBuffer({texture},{buffer,bytesPerRow:960*4},[960,100]);
+      GPU.device.queue.submit([encoder.finish()]);await buffer.mapAsync(GPUMapMode.READ);
+      const pixels=new Uint8Array(buffer.getMappedRange());let max=0;
+      for(let y=0;y<100;y++)for(let x=0;x<960;x+=2)for(let c=0;c<2;c++) {
+        const i=(y*960+x)*4+c;max=Math.max(max,Math.abs(pixels[i]-pixels[i+4]));
+      }
+      buffer.unmap();buffer.destroy();texture.destroy();return max;
+    }""")
+    assert noise<=1,noise
+    print('PASS smooth noise is continuous across horizontal and vertical lattice boundaries',noise)
+    motion=pg.evaluate("""async()=>{
+      const results=[],now=performance.now.bind(performance);performance.now=()=>5000;
+      try {
+        SET.fog=false;SET.glow='low';SET.lightEvery=1;R.lens=false;
+        for(const look of ['pixel','smooth'])for(const light of ['full','fast','off']) {
+          SET.look=look;SET.lighting=light;
+          for(let n=0;n<12;n++) {
+            SET.zoom=6-n*0.37;resizeRender();R.camX=80.25+n*7.3;R.camY=5.5+n*0.43;
+            R.rain=0;R.frame++;drawWorld();await GPU.device.queue.onSubmittedWorkDone();
+          }
+          const a=worldPixelContext().getImageData(0,0,960,600).data;
+          const cleanCells=GPU.uploadCells;
+          // Discard all view textures/light history as well as the resident-world cache.
+          GPU.resident.fill(0);GPU.viewKey=null;GPU.lightFrame=undefined;
+          R.rain=0;drawWorld();await GPU.device.queue.onSubmittedWorkDone();
+          const b=worldPixelContext().getImageData(0,0,960,600).data;
+          let max=0;for(let i=0;i<a.length;i++)max=Math.max(max,Math.abs(a[i]-b[i]));
+          results.push({look,light,max,cleanCells});
+        }
+        SET.lighting='full';return results;
+      } finally {performance.now=now;}
+    }""")
+    assert all(r['max']==0 for r in motion),motion
+    print('PASS pan/zoom matches fresh full redraw in both looks and all light modes',motion)
     cadence=pg.evaluate("""()=>{
       SET.lightEvery=3;R.frame=1002;drawWorld();const old=GPU.lightFrame,origin=R.lx0;
       R.camX+=4;R.frame=1003;drawWorld();const held=GPU.lightFrame===old && R.lx0===origin;

test/t_gpu_golden.py

@@ -1,4 +1,4 @@
-"""Frozen legacy references, captured before retirement; compare the WebGPU world only."""
+"""Frozen world references; seven smooth crops corrected for continuous noise in t91."""
 import base64, gzip, json, pathlib, sys, subprocess
 sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent.parent / 'tools'))
 from use_venv import ensure_playwright; ensure_playwright()