[Mocha] Fix RabbitMQ recovery, publishing and outbox dispatch - #10543

…hem away

The transport now owns connection recovery instead of the client's automatic
recovery, and re-attaches a consumer with backoff when the broker closes its
channel or cancels it while the connection stays up.

- Re-attach consumers whose channel was closed or whose subscription was
  cancelled, and declare the receive topology of every consumer before any
  consumer resumes on a connection.
- Dispose lost connections and reconnect in the background with backoff.
- Wait for the broker at startup for up to `ConnectTimeout`, two minutes by
  default, and fail at once on refused credentials, access or topology.
- Publish with confirms, return messages sent to a missing queue, retry
  publishes on connection loss, and cap the number of rented publish channels.
- Ack the redelivery of a message whose handler already completed without
  running the handler again.
- Let reply queues expire instead of being deleted with their last consumer,
  and keep bus queue defaults away from them.
- Declare exclusive queues only on the consumer connection.
- Name connections by role and require RabbitMQ.Client 7.2.2.

Enforce connection timeouts, cancel consumer attachment during shutdown, and retry losses that arrive as recovery completes. Reuse consumer workers across reconnects, cancel handlers on channel loss, and expose blocked recovery health. Remove the completed-delivery cache and cover recovery, capacity limits, redelivery, and shutdown with regression tests.

Preserve compatible queue defaults and explicit argument overrides while skipping quorum and stream defaults for non-durable queues. Cover both queue type configuration forms with regression tests.

Clarify that shutdown cancellation ends handler draining while channel cleanup continues independently.

Share publisher channels per endpoint and bound outstanding publish attempts while waiting for individual broker confirmations.

Add configurable Postgres outbox concurrency with independent workers so blocked dispatches do not stall idle capacity. Preserve cancellation behavior with explicit loop termination conditions.

Add concurrency and cancellation coverage and document the publishing and outbox settings.

Wait for consumer capacity before reattaching and for the local reply
consumer before dispatching requests. Cancel pending connection waiters
and drain connection setup during disposal, with idempotent consumer
shutdown and forced-stop escalation.

Keep request deadlines active until dispatch completes, including when a
response or fault arrives before dispatch finishes. Apply explicit quorum
replication defaults to generated reply queues so requests can recover
through surviving cluster nodes.

Respect broker defaults for unspecified queue types and delivery limits.
Honor explicitly configured limits with RabbitMQ 3 compatibility, and
preserve existing queues and their messages during provisioning.

Cover broker process crashes, repeated restarts, quorum leader failover,
connection failover, cancellation, disposal races, request deadlines, and
finite or unlimited redelivery with regression tests.

Validation: 223 targeted tests passed across RabbitMQ 3 and 4 and core
request deadlines. Documentation formatting and rendered pages checked.

Add correlated recovery, retry, fault, publish-wait, and drain diagnostics. Preserve connection and channel identity across recovery, clarify health failures, and correct conversation tracing.

Remove log-capture assertions and retain behavioral coverage for delivery, recovery, cancellation, and faults. Document scope export configuration.

Configure the cluster fixture to return disconnected-consumer deliveries after one second and verify the broker is running after each restart. Add coverage for an unacknowledged event held by a killed consumer node.

RabbitMQ transport

- Replace the connection manager base class with one internal
  RabbitMQConnectionManager: one connect routine with a deadline per
  attempt, a ready signal, and a recovery that only the first loss report
  starts. A connection attempt against an endpoint that never answers now
  ends.
- Publish on an exclusive channel lease from a pool per connection, with at
  most 100 channels. A channel is reused only after the broker gave a
  verdict on it, so a rejected or cancelled publish no longer affects other
  publishes.
- Add the PublishTimeout transport setting (default 2 minutes). A send or
  publish that the broker does not confirm in time fails with the new
  TransportUnavailableException.
- Keep the connection during a broker memory or disk alarm and do not send
  waiting messages again.
- Run each consumer as one lifecycle loop. Cancelling or stopping one
  consumer no longer aborts the shared connection, and a consumer whose
  handlers block reports an error only after a grace period.
- Remove the wait of receive endpoints for the reply endpoint.
- Classify failures as transient or permanent. A message the broker can
  never accept is no longer retried without end.
- Rename the log events and log one event per fact.
- RabbitMQDispatcher, RabbitMQConsumerManager and
  RabbitMQConnectionManagerBase are no longer public.

Outbox

- Dispatch with independent workers. A failing dispatch no longer cancels
  the others, which sent an already sent message again.
- Add DispatchTimeout, MaxAttempts and MaxRetryDelay. An unavailable
  transport does not use up attempts, and an exhausted message is logged as
  Critical with its id, type and destination.

Core and other transports

- Stopping with a cancelled token still cleans up in the Postgres and Azure
  Service Bus transports.
- A failed transport start stops what already started.
- The message bus health check takes a name, failure status, tags and a
  timeout, and uses the ready tag by default.
- The transport circuit breaker logs under its own category.
- ExponentialBackoff applies jitter at the cap as well.

Tests and docs

- Recovery tests prove the loss they recover from, with shared broker fault
  helpers and one timeout policy. Mock-based tests are replaced by tests
  against a broker.
- Document recovery, the publish timeout, broker alarms, the outbox options
  and the health check.