ci: Grid hypothesis falsified; stall probe instrumentation; honest ciSkips

The ?nogrid experiment disproved the Grid theory: rtt/tsl-halftone stall
grid-less and sprites (never had a grid) stalls too, while shadow-contact
(custom nodes + useUniforms) passes. Four stalls -> neutral ciSkip; full
pass/fail matrix + falsified/open hypotheses recorded in HANDOFF.

ReadinessSignal now publishes __frameCount/__loadersActive and the smoke
timeout message reports them - the next red run classifies the stall type
(dead loop / recompile crawl / loaders never settle) instead of guessing.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: DennisSmolek

docs/HANDOFF.md

@@ -33,14 +33,24 @@ cheaper as AGENTS.md absorbed each round's lessons (now at v0.4, see its changel
   **the research-designed WebGPU path is proven on free runners**.
 - `packageManager` pin + vendored fiber tarball (1.3MB, UPSTREAM.md A1) were needed
   to make CI installable.
-- **SwiftShader stall pattern (open investigation)**: examples combining drei's Grid
-  WITH a custom node graph (render pipeline or custom outputNode) hang readiness
-  silently — zero page errors, 2×180s. Grid-only passes; custom-nodes-only passes.
-  Affected: skinning-instancing, rtt, tsl-halftone → `ciSkip` in the manifest
-  (exception list per SPEC §10, each with the reason). All three pass on Metal/real
-  GPUs. Bisect idea: CI matrix job rendering rtt with grid off vs on. Possibly a
-  drei-Grid-shader trigger (fwidth/discard under SwiftShader) — could merge with
-  UPSTREAM B6 once bisected.
+- **SwiftShader stall (open investigation — Grid hypothesis FALSIFIED)**: four
+  examples hang readiness silently on SwiftShader (zero page errors, 2×180s); all
+  pass on Metal. The `?nogrid` experiment disproved the Grid theory (rtt/halftone
+  still stall grid-less; sprites stalls with no grid at all). Full matrix:
+  - STALL: skinning-instancing, rtt, tsl-halftone, sprites
+  - PASS: animation-skinning-blending (9s), hello-webgpu (5s),
+    postprocessing-bloom-emissive (24s), sky (7s), shadow-contact (7s)
+  - Not yet separated by: render pipeline (bloom passes, rtt stalls), Grid
+    (falsified), fiber `useUniforms` (shadow-contact calls it and passes),
+    animation (anim-blending passes).
+  - Instrumentation added: readiness timeouts now report `__frameCount` /
+    `__loadersActive` in the failure message — next red run classifies the stall
+    (0 frames = dead loop; few = per-frame pipeline recompile crawl; many =
+    loaders never settle). Bisect from that data.
+  - Mechanisms: `ciSkip` (skip with reason) and `?nogrid`/`ciNoGrid` (run grid-less)
+    both exist in the manifest + smoke spec; the four stalls currently use `ciSkip`.
+  - Smoke job is `continue-on-error` (advisory) until this is resolved — no failure
+    emails; flip back in ci.yml when stable.
 
 ## For Dennis
 

src/examples.json

@@ -17,7 +17,7 @@
     "tags": ["animation", "skinning", "instancing", "gltf", "tsl", "post-processing", "leva"],
     "original": "https://threejs.org/examples/#webgpu_skinning_instancing",
     "credits": "Michelle model from the three.js examples",
-    "ciNoGrid": "SwiftShader stall when Grid + custom node graph coexist (see HANDOFF); grid suppressed via ?nogrid in CI so the example itself still smoke-tests"
+    "ciSkip": "SwiftShader readiness stall (passes on Metal/real GPUs) — cause under investigation, see HANDOFF stall matrix"
   },
   {
     "slug": "postprocessing-bloom-emissive",
@@ -39,7 +39,7 @@
     "tags": ["tsl", "render-to-texture", "post-processing", "pointer"],
     "original": "https://threejs.org/examples/#webgpu_rtt",
     "credits": "uv_grid_opengl texture from the three.js examples",
-    "ciNoGrid": "SwiftShader stall when Grid + custom node graph coexist (see HANDOFF); grid suppressed via ?nogrid in CI so the example itself still smoke-tests"
+    "ciSkip": "SwiftShader readiness stall (passes on Metal/real GPUs) — cause under investigation, see HANDOFF stall matrix"
   },
   {
     "slug": "shadow-contact",
@@ -53,13 +53,14 @@
     "tags": ["tsl", "node-material", "gltf", "leva"],
     "original": "https://threejs.org/examples/#webgpu_tsl_halftone",
     "credits": "Michelle model from the three.js examples",
-    "ciNoGrid": "SwiftShader stall when Grid + custom node graph coexist (see HANDOFF); grid suppressed via ?nogrid in CI so the example itself still smoke-tests"
+    "ciSkip": "SwiftShader readiness stall (passes on Metal/real GPUs) — cause under investigation, see HANDOFF stall matrix"
   },
   {
     "slug": "sprites",
     "title": "Sprites",
     "tags": ["tsl", "node-material", "sprites", "fog", "leva"],
     "original": "https://threejs.org/examples/#webgpu_sprites",
-    "credits": "sprite1.png texture from the three.js examples"
+    "credits": "sprite1.png texture from the three.js examples",
+    "ciSkip": "SwiftShader readiness stall (passes on Metal/real GPUs) — cause under investigation, see HANDOFF stall matrix"
   }
 ]

src/utils/ReadinessSignal.tsx

@@ -14,6 +14,10 @@ import { useProgress } from '@react-three/drei/core'
 declare global {
   interface Window {
     __exampleReady?: boolean
+    /** Diagnostics for CI stall triage: total finish-phase frames since mount. */
+    __frameCount?: number
+    /** Diagnostics: last-seen loader activity (drei useProgress). */
+    __loadersActive?: boolean
   }
 }
 
@@ -26,12 +30,15 @@ export function ReadinessSignal() {
 
   useEffect(() => {
     window.__exampleReady = false
+    window.__frameCount = 0
     return () => {
       window.__exampleReady = false
     }
   }, [])
 
   useFrame(() => {
+    window.__frameCount = (window.__frameCount ?? 0) + 1
+    window.__loadersActive = useProgress.getState().active
     if (useProgress.getState().active) {
       settled.current = 0
       window.__exampleReady = false

tests/smoke.spec.ts

@@ -40,10 +40,19 @@ for (const { slug, ...meta } of examples) {
         timeout: process.env.CI ? 180_000 : 60_000,
       })
     } catch (cause) {
-      // Readiness timeouts are usually a dead render loop, not slowness — surface
-      // everything the page said so CI logs are diagnosable without an artifact dig.
+      // Surface everything the page can tell us so CI logs classify the stall:
+      // frames=0 → render loop never ran; frames tiny → crawl (per-frame pipeline
+      // recompile?); frames huge → loaders never settled (readiness logic).
+      const probe = await page
+        .evaluate(() => ({
+          frames: window.__frameCount ?? -1,
+          loadersActive: window.__loadersActive ?? null,
+          ready: window.__exampleReady ?? null,
+        }))
+        .catch(() => null)
       throw new Error(
-        `readiness timeout for ${slug}; page reported ${errors.length} error(s):\n` +
+        `readiness timeout for ${slug}; probe=${JSON.stringify(probe)}; ` +
+          `page reported ${errors.length} error(s):\n` +
           (errors.join('\n') || '(no console/page errors captured)'),
         { cause },
       )