A client connection that read a full response while the request body was
still flushing would see incorrect behavior, since the pool would let it
be checked out again for a new request. In debug builds, it would then
panic, but in release builds it would intermix the 2nd request bytes
with the body of the previous request.
In practice, this only ever happens if a server replies with a full
response before reading the full request, while also choosing to not
close that connection. Most servers either wait for the full request, or
close the connection after the new response is written, so as to stop
reading.
c88011da4e