[luke-lombardi]

Preserve guest OOM counters and init exit status in final metrics so fast exits retain their real cause. Counters reset on restore.

Validated native tests and 15 staging suites covering exits/signals, OOMs, sandbox survival, logging, functions/endpoints, cooldown, and checkpoint compatibility.

[jackf-beamcloud]

One real issue: the /task/init_exit_status callback calls l.k.GlobalInit().ExitStatus() unconditionally, but globalInit is nil until the root process is created, so any metrics snapshot taken between RegisterExitMetrics() and CreateProcess (metric server scrape / MetricsExport RPC while the sandbox is created but not started) will nil-deref inside ExitStatus and panic the sentry. Guard it (return 0 if GlobalInit() == nil). Non-blocking: the value is 0 while init is still running, indistinguishable from a clean exit, so consumers should only trust it from the final metrics file; and both OOM counters share the same Description, worth making guest_init_oom_kills say it's init-only. OOM counting in killMemoryOOMVictim and the tests look right.

[luke-lombardi]

Fixed in 3bd9d11: guard absent init and give the init-only OOM counter its own description. Added a real Metrics.Export RPC before container start; it fails without the guard and passes with it. Native OOM/final-exit tests pass. Beta9 reads this gauge only from the final metrics file after wait.

[jackf-beamcloud]

3bd9d11 fixes the nil-init deref (returns 0 before GlobalInit exists), gives guest_init_oom_kills its own description, and the pre-start Metrics.Export in the test covers the crash path. With beta9 only reading init_exit_status from the final metrics file, the 0-while-running ambiguity is fine. LGTM.