v1.4.1 — ai-service without RabbitMQ, undelivered AI requests reported

One branch, fix/ai-start-without-rabbitmq: the AI assistant no longer takes the rest of
ai-service down with RabbitMQ, and an AI request that cannot be delivered is reported instead of
lost. Found while production's CloudAMQP vhost was missing after 1.4.0 shipped.

ai-service starts without RabbitMQ

• ai-service serves HTTP while the broker is unreachable. It used to wait for the ai.requests
  consumer before listening, so with RabbitMQ down /health, the usage and quota screens, AI
  sessions and the confirm / cancel buttons of action cards all answered 502, and the Mac mini
  redeploy rolled every deploy back. The consumer now attaches by itself whenever the broker
  connects — at boot, after an outage, and after every reconnect — with no restart. The library
  logs "Skipping RabbitMQ Handlers"; Consuming ai.requests confirms the subscription.

AI requests that cannot be delivered

• The asker hears at once that the assistant is unavailable. When chat-service could not hand
  an AI request to ai-service (RabbitMQ down, history lookup failed) the error was swallowed: no
  log line, and the asker watched a reply bubble that never answered. Now chat-service logs a
  warning and sends AI_STREAM_ERROR / AI_UNAVAILABLE on the conversation's AI stream, which web
  and mobile already show as "AI is temporarily unavailable".

Versions: 1.4.1 (web, NestJS services), 1.4.1+6 (mobile, unchanged apart from the version).