mirror of
https://github.com/stablyai/orca.git
synced 2026-10-23 08:03:39 +00:00
Jinwoo-H
d7fdccf082
Two reviewers landed on the same remaining hole from opposite directions: the identity check and the unlink were still two operations, so a relay that bound the path in between was the one displaced. Adding a host-side lock would have to be shared with relay startup to mean anything. The unlink was never ours to do. A fresh daemon already reclaims a stale path itself, checking the socket's identity and binding in one process — and if it loses that race it fails to start rather than stranding anyone. So deploy now only asks who owns the path, and launches when the answer is that nothing does. The whole check-to-unlink window goes away, along with the identity bookkeeping that existed to narrow it. Proven by the journey where a killed relay's leftover socket is taken over by the production launch path. Reset drops the same distinction it could not support: an attempted signal is not evidence a process took it, so with no inventory to ask afterwards the stop is reported unproven whether or not something was signalled. The reset sandbox's lsof stub now answers the tool-works probe independently of the socket query — tying them together made a working lsof look broken the moment the relay stopped, which is not how a host behaves.