Attribute performance reports via JS self-profiling and denoise long-task prompts / 性能报告接入采样归因并为长任务提示降噪

[SivanCola]

Problem

Performance-pressure diagnostics dominate crash triage volume (~18k reports over 30 days across performance.longtask / performance.lag / performance.heap, versus 400 for the highest-volume real crash group), yet none of them carry a stack — longtask PerformanceObserver entries have no attribution beyond a culprit descriptor, so the reports are unactionable. Separately, the cumulative long-task rule (>=3 tasks && >=1.5s total per 60s) fires on routine streaming-render bursts (e.g. 16 tasks of 70–240ms), which is a large part of that volume.

Changes

Attribution (frontend + Go asset server)

• Run a rolling JS Self-Profiling sampler (10ms interval, 60s buffer) while the performance monitor is installed. When a performance prompt fires, stop the profiler, keep only samples that landed inside recorded long-task windows (plus the lag spike window itself for event-loop reports), aggregate leaf frames by self-time, and append a long task top frames (sampled) section to the report, e.g. 42x post (wails://wails/assets/vendor-markdown.js:1:130216).
• Serve Document-Policy: js-profiling from the Wails asset server (jsProfilingMiddleware, chained before the workspace-media middleware). Chromium WebViews (WebView2 — the dominant platform for these reports) honor it; WebKit ignores both the header and the API, so macOS/Linux keep today's report shape.
• Record longtask culprit attribution on each sample, but only surface anomalies (cross-context culprits / named containers); self/window carries no signal and is omitted.
• The report privacy note now mentions sampled app function names (en / zh / zh-TW).

Denoise

• Raise the cumulative long-task prompt budget from 1.5s to 3s per 60s window (single-task threshold stays 800ms). The trigger is extracted as shouldPromptForLongTasks with direct test coverage.

Backward compatibility

Surface	Old-data behavior	Conclusion
CrashPayload (schemaVersion: 2)	Unchanged fields; new content is additive lines inside the freeform message string	Safe — server/message consumers see the same schema
PerformanceSnapshot	In-memory only, never persisted; new optional fields (attribution, longTaskFrames)	Safe
localStorage reasonix:perf-reported	Format untouched	Safe
Asset responses	New Document-Policy header; WebKit/older WebViews ignore unknown policies	Safe
Profiler API absent (WebKit, older WebView2)	startLongTaskProfiler no-ops / catches, prompt path falls back to the synchronous unattributed report	Safe

Verification

• tsx src/__tests__/crash-reporting.test.ts — 60 passed, 0 failed (new coverage: threshold boundaries incl. the 16×~100ms streaming burst staying silent, attribution formatting, trace aggregation windowing/capping, report context formatting)
• pnpm test (full frontend suite) — pass
• tsc --noEmit -p tsconfig.test.json — pass
• cd desktop && go build ./... && go test . — pass (rebased onto latest main-v2, which also touched app.go and the locale files)

Follow-up (separate PR): the sampled data is expected to confirm full-document markdown re-parse per streaming frame as the dominant long-task source; block-level memoization of completed markdown blocks is the planned fix.