feat(observability): correlate agent relay delivery

[MajorTal]

Trying to debug why my agents sometimes do not respond, I realised the current diagnostics are insufficient. Which is why this PR adds a focused correlation trail across the relay, Redis pub/sub, and the ACP harness.

Today a delayed response can look like a generic WebSocket stream ending followed by a later reconnect. The logs do not preserve the Close code/reason, join the client socket to the server-side connection, or show whether Redis and local WebSocket fan-out handled the event live.

What changed

• Return an opaque x-buzz-connection-id on the WebSocket upgrade so ACP and relay logs identify the same socket.
• Record ACP connection generations, Close code/reason, error/EOF/ping-timeout exits, and event receipt age/replay hints.
• Record Redis publish subscriber counts and relay fan-out recipient/drop counts using the existing event ID.
• Add Redis subscriber connection, disconnect, reconnect, and topic gauges, and enable buzz_pubsub=info in the relay's default log/OTEL filters.

The new diagnostics contain IDs, timestamps, and counts only—no message contents, auth material, or private keys. Delivery behavior is otherwise unchanged.

What this should answer next time

• Did the ACP socket close cleanly, time out, reset, or disappear without a Close frame?
• Which relay-side connection handled that exact socket?
• Did Redis report any subscribers when the event was published?
• Did the receiving relay match any local WebSocket subscriptions?
• Did ACP receive the event live or only after reconnect/replay?

Validation

• cargo fmt --all --check
• cargo clippy -p buzz-acp -p buzz-pubsub -p buzz-relay --all-targets -- -D warnings
• cargo test -p buzz-acp --lib — 669 passed
• cargo test -p buzz-pubsub --lib — 24 passed, 11 ignored
• Targeted relay and diagnostics tests pass
• cargo test -p buzz-relay --lib — 839 passed, 37 ignored; demo_join_forwarded_arm_round_trips_echo fails with the same 504 vs 200 assertion on this branch and an untouched detached origin/main worktree

[MajorTal]

Filed #5472 for the diagnostic gap this addresses, per CONTRIBUTING.md's issue-first guidance. The open judgment call is log volume — making this useful by default means enabling buzz_pubsub=info in the relay's default filters. That trade is better decided there than in a diff review.