relay host connection: bound retained failed channels; report a failed record refresh; distinct test titles #783

Follow-up from the #759 review (non-blocking, LOW).

What the fix does:

1. Bound the retained failed channels. The relay host connection keeps channels whose close() rejected so stop() can close them later. It keeps at most a fixed number (or the latest one plus a count), so a long outage on a transport whose close keeps failing cannot turn one stop into many serial close timeouts or hold many sockets.
2. Report a failed record refresh. The heartbeat and health callbacks persist the host record inside a try/catch that discards the error, and since fix(cli): serialize office kill and the host connection writer per host record #759 the persist can fail on a held host lock. Log the first failure per connection, naming the record and the reason, so a stale record has a visible cause.
3. Distinct test titles. test/relay-host-record-719.test.ts names its two close rejects cases with %s over booleans, which prints literally under Bun 1.4.4. Put the value in the title so the two cases are told apart in the report.

Each change comes with a test (a mutation that removes it turns the test red).

Refs #748, #759.