stdio client keeps dialing a dead port after domain reload — cache TTL + backoff ignore the status file the editor already wrote

[BatuBoyGG]

In batch-mode sessions, every script-recompiling refresh triggers a domain reload that kills the stdio bridge listener. The editor side handles this well: StdioBridgeHost retries the old port for BusyPortFallbackWindowSeconds (3 s), can't rebind because the socket is in TIME_WAIT (~30 s on macOS, and ExclusiveAddressUse = true is deliberate per #1173), walks to the next port, and immediately announces the new port via the status file — exactly as its own warning log says: "clients follow via the status file."

The Python server doesn't follow promptly. Two compounding delays:

1. port_registry_ttl (5.0, core/config.py, no override available) — after the reload, the client keeps dialing the dead cached port until the TTL expires, even though the status file already names the live port.
2. Exponential reconnect backoff — by the time the editor is back (~4 s of real reload work), the client is sleeping and overshoots.

Measured on macOS / Unity 2022.3.62f2 / mcpforunityserver 10.1.2, batch mode: ~12–15 s of reconnect latency per reload against ~4 s of actual reload work, deterministically on every reload (an automated pipeline of ours pays it 1–3 times per job; the editor walked 6400 → 6401 → 6402 across two reloads in one session, recovery always correct, just slow).

Proposed fix

On connection refused/reset, treat the cached discovery as invalid immediately — re-read the status registry before entering backoff. The fresh port is already on disk at that moment; the current code just doesn't look.

Alternatively (or additionally): make port_registry_ttl and BusyPortFallbackWindowSeconds configurable via environment variables so automated hosts can tune them.

Happy to provide editor logs / timings from the runs where we measured this.

[SeojunKim-pumisj]

I’d like to work on this. I’ll update the stdio reconnect path to bypass the discovery TTL after a communication failure, and add a regression test verifying the refreshed registry port is used before retry backoff.