GeezerVison

Under highly contentious load, there might be a race condition in subscription/listen.
Every change broadcast before subscribing to the change feed is lost.
Subscription happens on a thread, and writing to the wire happens in another thread.

The `ack` might be fully written to the transport, and received by the client,
before the changes are subscribed to.
A test harness might fire a change right after the test client receives `ack`,
but before the changes are subscribed to, and therefore the change is lost.

So we need to ensure that:
1. Changes are subscribed to BEFORE the ack
2. But the ACK goes first in any case (and only once)

This is mostly a test-harness fix rather than a real-world scenario.

Signed-off-by: Daniel Garnier-Moiroux <[email protected]>