attach() now returns False both when the client dropped mid-replay (session
already detached, socket dead) and when the force-redraw write stalled
(socket still attached, worth closing with 1013). Only close the socket in
the second case; the registry detach stays as the guard for both.
Also trims the salvaged tests to two invariants, both red on origin/main:
a mid-replay drop leaves the session reap_idle()-reclaimable, and a drain
send failure detaches the current socket without touching a replacement
that attached during the send (#110849).
teknium1
7a1d278528