[remotego]

Summary

The router boot script from #72 runs a silent gap between the argv line (router: prefill ... / router: decode ...) and the router process actually starting — during which probe_identity may be querying sidecars, writing the snapshot file, or (worst case) stuck on the 90-minute budget. kl logs on a router pod showed nothing between those two moments, which made it hard to tell whether boot was progressing or wedged.

Three breadcrumbs added, all on stderr so probe_identity's stdout return value stays clean:

router: <ip> identity <restarts>|<watcher_started_at>   # per-IP success
router: <ip> identity sentinel (...reason...)           # per-IP sentinel (old sidecar / hw disabled)
router: boot script done, exec into router              # handoff to exec

Sentinel reasons disambiguate the three no-identity paths:

• hangWatcher disabled at render — HW_PORT=0
• sidecar has no /engine — 404 from an older sidecar
• The existing 90min budget exhausted line (unchanged) covers the third case.

Rendered-output change

Only the boot-script body changes — adds the echo >&2 lines, no behavior change. Affects both charts' pd-ds DisaggregatedSet router Deployment. 52 non-pd cases unchanged.

Gold rebaked per R3 at this branch's HEAD; 54/54 verify.

Test plan

• helm template renders clean for all pd cases, both charts
• bash -n on the extracted boot scripts
• E2E against mock hang-watchers: 200 → stdout carries tuple, stderr carries log line; 404 → sentinel on stdout, reason on stderr; hw-disabled → same shape
• Live deploy: kl logs <router-pod> on a pd render shows the three breadcrumbs in order (argv line → per-IP identity/sentinel → handoff → router's own first log)

🤖 Generated with Claude Code