fix(engine): bucket gpu_renderer + cover the failure cohort (review)

Three review findings on the win32 drawElement PR:

1. gpu_renderer shipped the raw UNMASKED_RENDERER_WEBGL string — unbounded,
   driver-authored, GPU-model-specific, and |-joined across parallel
   sessions, i.e. high cardinality by construction, against this file's own
   convention of sanitizing engine-sourced strings (deGateReason is a
   bucket; error messages go through redactTelemetryString). Now bucketed at
   the source by classifyGpuRenderer to <backend>/<vendor>
   (metal/apple, d3d11/nvidia, swiftshader/other, ...), which is the whole
   analytic signal the win32 rollout needs and nothing else. The raw string
   never leaves the engine.

2. gpu_renderer reached render_complete only, so a crashed render — the
   cohort the field exists to attribute — carried no backend. It now rides
   RenderCaptureObservability (deGpuRenderer, sourced from the live probe
   session like the de_* counters), so both render_complete and
   render_error carry it and a hard failure still reports its GPU backend.
   On render_complete the perfSummary value still wins by spread order.

3. Restore the fallow-ignore-next-line suppression above
   __resetDeParallelRouterTrialStateForTests: CLI test files are not fallow
   entry points, so removing it fails the CI dead-code audit (local
   pre-commit passed only because of its changed-file scope).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

packages/engine/src/services/drawElementService.test.ts

@@ -1,6 +1,7 @@
 import { describe, expect, it, vi } from "vitest";
 import type { Page } from "puppeteer-core";
 import {
+  classifyGpuRenderer,
   detectGpuBackend,
   detectSwiftShader,
   resolveDrawElementCaptureMode,
@@ -48,6 +49,42 @@ describe("detectGpuBackend", () => {
   });
 });
 
+// ── classifyGpuRenderer ────────────────────────────────────────────────────────
+
+describe("classifyGpuRenderer", () => {
+  it("buckets real ANGLE renderer strings to <backend>/<vendor>", () => {
+    expect(
+      classifyGpuRenderer("ANGLE (Apple, ANGLE Metal Renderer: Apple M4 Pro, Unspecified Version)"),
+    ).toBe("metal/apple");
+    expect(
+      classifyGpuRenderer(
+        "ANGLE (NVIDIA, NVIDIA GeForce RTX 3080 Direct3D11 vs_5_0 ps_5_0, D3D11)",
+      ),
+    ).toBe("d3d11/nvidia");
+    expect(
+      classifyGpuRenderer("ANGLE (Intel, Intel(R) UHD Graphics 630 Direct3D11 vs_5_0 ps_5_0)"),
+    ).toBe("d3d11/intel");
+    expect(classifyGpuRenderer("ANGLE (AMD, AMD Radeon RX 6800 Direct3D11 vs_5_0 ps_5_0)")).toBe(
+      "d3d11/amd",
+    );
+    expect(classifyGpuRenderer("Google SwiftShader")).toBe("swiftshader/other");
+  });
+
+  it("drops the GPU model — the bucket must stay low cardinality", () => {
+    // Two different NVIDIA cards must collapse to ONE bucket, otherwise the
+    // property is unbounded and useless for aggregation.
+    expect(classifyGpuRenderer("ANGLE (NVIDIA, NVIDIA GeForce RTX 4090 Direct3D11)")).toBe(
+      classifyGpuRenderer("ANGLE (NVIDIA, NVIDIA GeForce GTX 1060 Direct3D11)"),
+    );
+  });
+
+  it("returns undefined for missing input rather than a bogus bucket", () => {
+    expect(classifyGpuRenderer(null)).toBeUndefined();
+    expect(classifyGpuRenderer(undefined)).toBeUndefined();
+    expect(classifyGpuRenderer("")).toBeUndefined();
+  });
+});
+
 // ── resolveDrawElementCaptureMode ──────────────────────────────────────────────
 
 describe("resolveDrawElementCaptureMode", () => {

packages/engine/src/services/drawElementService.ts

@@ -110,15 +110,58 @@ export interface GpuBackendInfo {
   isSwiftShader: boolean;
   /**
    * Raw UNMASKED_RENDERER_WEBGL string (e.g. "ANGLE (Apple, ANGLE Metal
-   * Renderer: Apple M4 Pro, ...)", "ANGLE (NVIDIA, D3D11 ...)"), or null when
-   * WebGL / the debug extension is unavailable. Carried to render telemetry:
-   * drawElement failure modes proved compositor-backend-specific during the
-   * macOS rollout, so the win32/D3D11 cohort needs damage clusters
-   * attributable to a GPU vendor + ANGLE backend, not just `os`.
+   * Renderer: Apple M4 Pro, ...)", "ANGLE (NVIDIA, GeForce RTX 3080 Direct3D11
+   * vs_5_0 ps_5_0, D3D11)"), or null when WebGL / the debug extension is
+   * unavailable. LOCAL USE ONLY — this is unbounded driver-supplied text and
+   * must not be shipped to telemetry verbatim; send
+   * {@link classifyGpuRenderer}'s bucket instead.
    */
   renderer: string | null;
 }
 
+/**
+ * Low-cardinality bucket for a raw WebGL renderer string: `<backend>/<vendor>`
+ * (e.g. `metal/apple`, `d3d11/nvidia`, `swiftshader/other`).
+ *
+ * drawElement failure modes proved compositor-backend-specific during the
+ * macOS rollout, so the win32/D3D11 cohort needs damage attributable to an
+ * ANGLE backend + GPU vendor. The raw string can't do that job in telemetry:
+ * it is unbounded, driver-authored, carries specific GPU model names, and is
+ * joined across parallel sessions — high cardinality by construction. The
+ * bucket keeps the analytic signal (which backend, which vendor) and drops
+ * everything else, matching how `deGateReason` is a sanitized bucket rather
+ * than the full fallback trigger. Pure; exported for tests.
+ */
+export function classifyGpuRenderer(renderer: string | null | undefined): string | undefined {
+  if (!renderer) return undefined;
+  const r = renderer.toLowerCase();
+  const backend = r.includes("swiftshader")
+    ? "swiftshader"
+    : r.includes("metal")
+      ? "metal"
+      : r.includes("direct3d11") || r.includes("d3d11")
+        ? "d3d11"
+        : r.includes("direct3d9") || r.includes("d3d9")
+          ? "d3d9"
+          : r.includes("vulkan")
+            ? "vulkan"
+            : r.includes("opengl") || r.includes("angle")
+              ? "opengl"
+              : "other";
+  const vendor = r.includes("apple")
+    ? "apple"
+    : r.includes("nvidia")
+      ? "nvidia"
+      : r.includes("amd") || r.includes("radeon")
+        ? "amd"
+        : r.includes("intel")
+          ? "intel"
+          : r.includes("microsoft")
+            ? "microsoft"
+            : "other";
+  return `${backend}/${vendor}`;
+}
+
 /**
  * Detect the page's WebGL backend: SwiftShader vs a real GPU, plus the raw
  * renderer string for telemetry.

packages/engine/src/services/frameCapture.ts

@@ -34,6 +34,7 @@ import {
   shouldDefaultCaptureBeyondViewport,
 } from "./screenshotService.js";
 import {
+  classifyGpuRenderer,
   detectGpuBackend,
   injectDrawElementCanvas,
   captureDrawElementFrame,
@@ -147,10 +148,12 @@ export interface CaptureSession {
   /** True if running on SwiftShader (detected at init). Undefined before init. */
   isSwiftShader?: boolean;
   /**
-   * Raw WebGL UNMASKED_RENDERER_WEBGL string, captured alongside the
-   * SwiftShader probe at DE session init (e.g. "ANGLE (NVIDIA, D3D11 ...)").
-   * Surfaces in CapturePerfSummary → render telemetry so backend-specific
-   * drawElement damage (Metal vs D3D11 vs GL) clusters attributably.
+   * Low-cardinality GPU bucket (`<backend>/<vendor>`, e.g. `d3d11/nvidia`)
+   * derived from the WebGL renderer at DE session init. Surfaces in
+   * CapturePerfSummary → render telemetry so backend-specific drawElement
+   * damage (Metal vs D3D11 vs GL) clusters attributably. The raw
+   * driver-supplied string is deliberately NOT retained — see
+   * classifyGpuRenderer.
    */
   gpuRenderer?: string;
   /** drawElementImage canvas was injected and is ready for capture. */
@@ -713,7 +716,7 @@ async function initDrawElementOrTransparentBackground(
   if (useDrawElement) {
     const gpuBackend = await detectGpuBackend(page);
     session.isSwiftShader = gpuBackend.isSwiftShader;
-    session.gpuRenderer = gpuBackend.renderer ?? undefined;
+    session.gpuRenderer = classifyGpuRenderer(gpuBackend.renderer);
     const transparent = session.options.format === "png";
     async function routeToFallback(): Promise<void> {
       session.captureMode = session.launchCaptureMode;