fix(gpu): centralize process type detection and harden marker durability

- Extract isGpuProcessType to shared module to prevent classification drift across crash reporting, fallback budget, and dialog logic
- Add fsync() before marker rename to survive power loss during persistence
- Resolve GPU fallback environment before recording breadcrumb, so triage doesn't see an engaged fallback that never ran
- Derive MIN/MAX tier constants from GPU_FALLBACK_TIERS array to keep clamping in sync if the ladder changes

src/main/crash-reporting/gpu-crash-fallback-decision.ts

@@ -1,3 +1,5 @@
+import { isGpuProcessType } from '../../shared/crash-reporting'
+
 export type GpuCrashFallbackOptions = {
   /** Rolling span over which clustered GPU crashes indicate a broken driver. */
   windowMs: number
@@ -86,7 +88,7 @@ export class GpuCrashFallbackTracker {
 
 /** True for the Chromium child process types whose crashes should count here. */
 export function isGpuChildProcessType(processType: string | undefined): boolean {
-  return (processType ?? '').toLowerCase() === 'gpu'
+  return isGpuProcessType(processType)
 }
 
 export function isGpuFallbackCrashCandidate({

src/main/crash-reporting/process-gone-classification.ts

@@ -1,3 +1,5 @@
+import { isGpuProcessType } from '../../shared/crash-reporting'
+
 export type ProcessGoneSource = 'renderer' | 'child'
 export type ExpectedTeardownScope = 'none' | 'renderer-reload' | 'app-shutdown'
 
@@ -48,7 +50,7 @@ function isRecoverableChromiumChildProcess({
     // still tears the GPU child down noisily, so app-shutdown stays suppressed.
     return !(
       gpuFallbackActive === true &&
-      normalizedProcessType === 'gpu' &&
+      isGpuProcessType(processType) &&
       expectedTeardown !== 'app-shutdown'
     )
   }

src/main/crash-reporting/process-gone-recorder.ts

@@ -2,6 +2,7 @@ import os from 'node:os'
 import { app } from 'electron'
 import {
   isCrashReportReason,
+  isGpuProcessType,
   sanitizeCrashReportString,
   type CrashReportBreadcrumbData
 } from '../../shared/crash-reporting'
@@ -60,7 +61,7 @@ function countsAgainstGpuFallbackReportBudget(event: ProcessGoneCrashEvent): boo
   return (
     event.gpuFallbackActive === true &&
     event.source === 'child' &&
-    event.processType.toLowerCase() === 'gpu'
+    isGpuProcessType(event.processType)
   )
 }
 
@@ -161,7 +162,7 @@ export function recordProcessGoneCrash(
   // Why: GPU and renderer deaths are the ones triage needs driver identity for;
   // every other child type would just pad the report.
   const gpuIdentity =
-    event.source === 'renderer' || event.processType.toLowerCase() === 'gpu'
+    event.source === 'renderer' || isGpuProcessType(event.processType)
       ? (getGpuInfoSnapshot() ?? {})
       : {}
   const crashDetails = buildProcessGoneCrashDetails({

src/main/index.ts

@@ -1651,6 +1651,12 @@ async function handleGpuChildCrash(reason: string, exitCode: number | null): Pro
     })
     return
   }
+  // Why: resolve first — a breadcrumb claiming the fallback engaged, written ahead of
+  // the guard that abandons it, reads in triage as a remediation that never ran.
+  const environment = getWindowsGpuFallbackEnvironment()
+  if (!environment) {
+    return
+  }
   recordCrashBreadcrumb('gpu_fallback_engaged', {
     reason,
     exitCode,
@@ -1682,10 +1688,6 @@ async function handleGpuChildCrash(reason: string, exitCode: number | null): Pro
     recordDurableCrashBreadcrumb('gpu_fallback_restart_deferred', fallbackData)
     return
   }
-  const environment = getWindowsGpuFallbackEnvironment()
-  if (!environment) {
-    return
-  }
   try {
     writeGpuFallbackMarker(
       userDataPath,
@@ -2147,12 +2149,13 @@ void app.whenReady().then(async () => {
   // Skipped in headless serve: no crash dialog to feed, and 'complete' can provoke a
   // GPU child that Xvfb environments have no reason to spawn.
   if (!isServeMode) {
-    void captureGpuIdentity(
-      (infoType) => app.getGPUInfo(infoType),
-      GPU_INFO_CAPTURE_TIMEOUT_MS
-    ).then((snapshot) => {
-      recordCrashBreadcrumb('gpu_info_captured', snapshot)
-    })
+    void captureGpuIdentity((infoType) => app.getGPUInfo(infoType), GPU_INFO_CAPTURE_TIMEOUT_MS)
+      .then((snapshot) => {
+        recordCrashBreadcrumb('gpu_info_captured', snapshot)
+      })
+      .catch(() => {
+        // best effort; GPU identity is a diagnostic, never a startup dependency
+      })
   }
   scheduleGpuFallbackHistoryReset()
   // Why: install certificate decisions before any webview or headless window issues its first TLS request.

src/main/startup/gpu-fallback-marker.ts

@@ -1,4 +1,12 @@
-import { readFileSync, renameSync, rmSync, writeFileSync } from 'node:fs'
+import {
+  closeSync,
+  fsyncSync,
+  openSync,
+  readFileSync,
+  renameSync,
+  rmSync,
+  writeFileSync
+} from 'node:fs'
 import { join } from 'node:path'
 import { clampGpuFallbackTier, type GpuFallbackTier } from './gpu-fallback-tiers'
 
@@ -137,7 +145,15 @@ export function writeGpuFallbackMarker(
   // reads as invalid and silently re-arms hardware acceleration next launch.
   const tempPath = join(userDataPath, GPU_FALLBACK_MARKER_TEMP_FILE)
   try {
-    writeFileSync(tempPath, JSON.stringify(marker))
+    // Why: a durable rename over non-durable data leaves an empty marker after power
+    // loss, which reads as invalid and re-arms hardware acceleration.
+    const handle = openSync(tempPath, 'w')
+    try {
+      writeFileSync(handle, JSON.stringify(marker))
+      fsyncSync(handle)
+    } finally {
+      closeSync(handle)
+    }
     renameSync(tempPath, markerPath(userDataPath))
   } catch (error) {
     try {

src/main/startup/gpu-fallback-tiers.ts

@@ -19,8 +19,11 @@ export const GPU_FALLBACK_TIERS = [1, 2] as const
 
 export type GpuFallbackTier = (typeof GPU_FALLBACK_TIERS)[number]
 
-export const MIN_GPU_FALLBACK_TIER: GpuFallbackTier = 1
-export const MAX_GPU_FALLBACK_TIER: GpuFallbackTier = 2
+// Why: derived, so adding a rung to the ladder cannot leave clamping and escalation pinned to the old ceiling.
+export const MIN_GPU_FALLBACK_TIER: GpuFallbackTier = GPU_FALLBACK_TIERS[0]
+export const MAX_GPU_FALLBACK_TIER: GpuFallbackTier = GPU_FALLBACK_TIERS.reduce((highest, tier) =>
+  tier > highest ? tier : highest
+)
 
 /** Tier 0 means "no fallback applied to this launch". */
 export const NO_GPU_FALLBACK_TIER = 0

src/shared/crash-reporting.ts

@@ -163,6 +163,11 @@ export function isReactErrorBoundaryReport(report: CrashReportRecord): boolean {
   )
 }
 
+/** Single definition of "this is the GPU child"; crash classification, budget accounting and the dialog must not drift apart. */
+export function isGpuProcessType(processType: string | undefined): boolean {
+  return processType?.toLowerCase() === 'gpu'
+}
+
 /**
  * A GPU child crash that happened while a GPU fallback tier was already applied.
  *
@@ -172,7 +177,7 @@ export function isReactErrorBoundaryReport(report: CrashReportRecord): boolean {
 export function isGpuFallbackCrashReport(report: CrashReportRecord): boolean {
   return (
     report.source === 'child' &&
-    report.processType.toLowerCase() === 'gpu' &&
+    isGpuProcessType(report.processType) &&
     typeof report.details.gpuFallbackTier === 'number' &&
     report.details.gpuFallbackTier > 0
   )