The synchronous drain loop introduced in 2a8718b correctly recovers
from exceptions, unlike the old promise chain which silently dropped
all subsequent messages after a throw. However, the recovery still
deferred remaining messages until the next onmessage event.

Wrap #processMessage in try/catch so each message is isolated — if one
fails (e.g. from a user callback), the rest continue immediately.

Also downgrade "unknown querySetId" logs from error to warn. These are
benign race conditions (e.g. server responding to a failed unsubscribe
after the subscription was already cleaned up) and not data corruption.