tenant.openagents/omega

Let the harness's reconnect actually run (omega#49)

Let the harness's reconnect actually run (omega#49) `c886698b35` added a reconnect-and-retry around the Sarah publish and it never executed once. Two device runs died the same way: device-proof: OWNER SEND ACCEPTED · message_seq 1 · turn.1 · run_state running panicked at sarah_conversation.rs:6342: NIP-42 authenticate: Relay("WebSocket protocol error: Connection reset without closing handshake") with `device-proof: sarah publish failed …; reconnecting` absent from both logs. The guard was at the wrong boundary. On a socket the relay has already reset, `publish` returns `IdentityRequired`, control enters the AUTH branch, and the AUTH write is what hits the reset — one level inside the function the retry wrapper wraps. An `expect` there takes the process down before any `Err` can propagate, so the wrapper could only ever have caught failures that did not happen. The AUTH path now returns instead of panicking, so the reconnect it was written for can run. A relay closing an idle connection is ordinary, and the harness reporting it as Sarah failing to answer is indistinguishable from the host never producing a reply — which is exactly how long this went unnoticed. The reply arm itself is proven: on a fresh socket the device rendered `Sarah · 1785040716 / Received. Answering turn.1 from a real Omega host over a real relay.` against the host's `SARAH REPLIED 7fb2f035c970e926 · turn.1`, not quarantined. That run only stayed up because the send landed ~15s after the socket opened. This removes the timing dependence. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>