[bvolpato]

Expansion service startup currently accepts any TCP listener on the selected port. Another process can claim that port before the Java jar binds, leaving the helper with an unrelated service address. In this CI run, the failed expansion address also appeared as a Python worker endpoint.

Check the expansion RPC before caching the address, detect an exited jar, and clean up failed startup. A wrong-service response or a jar that exits during startup gets a bounded retry on a fresh port. Expand-only services remain supported through a single probe within the startup deadline. The probe uses a fixed reconnect delay because the default gRPC backoff can outlast the startup deadline.

Reproduced the old helper accepting an unrelated gRPC listener and a real jar failing to bind the occupied port. Focused tests, the race detector, vet, and the Windows process-wrapper build passed. Real-jar startup, early exit, and startup-deadline cleanup were also checked. A local measurement with services that bound late in the startup deadline failed often with the default backoff and passed with the fixed delay. The full GoPortable container workflow was not run locally.