Report shared and service workers ending as page_worker_ended, not page_crashed

[archandatta]

Summary

• handleTargetCrashed now emits a new page_worker_ended event (category page, with target_id, target_type and url) for shared_worker and service_worker targets. Every other target type keeps emitting page_crashed.
• adds BrowserPageWorkerEndedEvent to openapi.yaml and to the KnownBrowserTelemetryEvent union, and regenerates oapi.go and category_gen.go
• adds a real-Chromium regression test, TestPageWorkerEndedChrome, and runs it in the CI "CDP telemetry browser regressions" step
• documents page_crashed and page_worker_ended in the cdpmonitor README

Why

The monitor mapped every Inspector.targetCrashed to page_crashed. Only three DevTools agent hosts in Chromium send that notification:

• RenderFrameDevToolsAgentHost::RenderProcessExited sends it only when the renderer terminates abnormally (crashed, killed or failed to launch). These are real crashes.
• SharedWorkerDevToolsAgentHost::WorkerDestroyed sends it every time a shared worker ends.
• ServiceWorkerDevToolsAgentHost::WorkerStopped sends it every time a service worker stops.

Dedicated workers, worklets, tabs and the browser target never send it. Because the monitor attaches to shared and service workers with Target.attachToTarget, sessions that never crashed reported a page_crashed each time one of those workers ended, which buried real renderer crashes.

Checked against real Chromium with the monitor attached:

scenario	before	after
shared worker calls self.close()	page_crashed	page_worker_ended
tab that owns a shared worker closes	page_crashed	page_worker_ended
service worker replaced by a newer version	page_crashed	page_worker_ended
ServiceWorker.stopAllWorkers	page_crashed	page_worker_ended
service worker left idle for 120 s	nothing	nothing
dedicated worker closes or terminates	nothing	nothing
renderer hosting a page and its shared worker is killed	page_crashed ×2	page_crashed (page) + page_worker_ended (worker)

Idle service workers are not stopped while a DevTools session is attached. ServiceWorkerVersion skips idle termination when devtools_attached() is true, and the monitor holds a session on every service worker. So idle timeouts do not produce this event; the schema and README say so.

CDP gives no way to tell a worker that ended normally from one whose process died, so both produce page_worker_ended. When a renderer dies, the pages it hosted still report page_crashed.

Unchanged

• The worker session stays tracked after page_worker_ended. A stopped service worker restarts on the same session and target, and the e2e test checks that its network telemetry keeps flowing and that its next stop reports the event again.
• page_crashed on page, iframe and background_page targets, and on untracked sessions (target_type: other), behaves as before. The page_crashed schema description notes that data from earlier images can still contain worker page_crashed events.
• OTLP severity: page_crashed stays ERROR and page_worker_ended gets the default INFO. Both are pinned in TestToLogRecord_Severity.

Generated code

The new schema pins its const names with x-enum-varnames. oapi-codegen v2's enum-collision pass compares names that may already carry a prefix, so adding another enum with the value page changes which existing enums get prefixed: without the pin, the existing unused oapi.Page const is renamed. With the pin, the regenerated oapi.go only adds code; the only removed line is the page_crashed description this PR edits. Regenerating unmodified main produces no diff.

The public API schema change is kernel/kernel#4615. The API passes unknown event types through on read, so the order in which the two land does not matter.

Testing

• go vet ./... — clean
• go test -race $(go list ./... | grep -v /e2e$) — all packages pass
• KERNEL_CDPMONITOR_CHROME_E2E=1 go test -race -count=1 ./lib/cdpmonitor — the full real-Chromium suite passes
• TestPageWorkerEndedChrome passed 10/10 under -race on Ubuntu Chromium 154 and 10/10 on Chrome for Testing 154.0.8037.92 (the version CI installs), using CI's exact -run filter.
• Mutation checks: routing worker targets back to page_crashed fails the shared and service worker subtests, and routing page and iframe targets to page_worker_ended fails the renderer-crash subtest plus TestTargetCrashed.
• The renderer-crash subtest SIGKILLs the renderer PIDs reported by SystemInfo.getProcessInfo. An earlier version used Page.crash, which goes through Chrome's crash handler; it passed locally but timed out once in CI, where findChromium may pick up a different binary.
• TestTargetCrashedByTargetType (fake CDP server) covers shared_worker and service_worker → page_worker_ended and iframe → page_crashed, and asserts that the session stays tracked.

🤖 Generated with Claude Code