A3 conformance probe can wedge relay boot without a deadline or early health listener

## Problem

At `7a9a523`, relay startup runs the A3 git object-store conformance probe as a hard boot gate. The default probe uses 32 writers across 3 rounds against blob storage. Individual requests have timeouts, but the probe has no overall deadline.

On a homelab MinIO deployment, the probe wedged silently for more than an hour during `if_none_match_race` round 0. The container remained `running` with zero restarts and emitted no further logs. Restarting the container fixed the issue; a normal probe completes in about five seconds.

The health listener on port 8080, including `/_liveness` and `/_readiness`, starts only after the boot gates complete. During the wedge, the Compose healthcheck received connection refusals until its full budget expired and marked the container unhealthy. Orchestrators cannot distinguish a relay that is still booting from one stuck in the probe.

## Suggested fixes

- Add an overall deadline to the A3 probe and fail startup with an explicit error when it expires.
- Bind the health listener before long-running boot gates so liveness and readiness remain observable during startup.

Read the relay startup path around the A3 git object-store conformance probe, then inspect the health listener endpoints `/_liveness` and `/_readiness` and the Compose healthcheck behavior. Done means the probe cannot wedge startup indefinitely, expiration reports an explicit error, and health remains observable while boot gates run.

Written by the indexing model from the issue text.