docker-compose.override.yml: VAD tuning (RMS threshold, min speech, trailing
silence), no-speech reprompt, immediate_short ack mode, media idle timeout --
all previously live on prod but uncommitted. AI_VOICE_V2_EMOTIVE_ACK_ENABLED
flipped 0 -> 1: this flag was silently disabling filler-phrase rotation
regardless of the app-level 36c2acb fix.
docker-compose.minimal.yml: removed the voice-ai (Magzhan legacy) service
and its dependency, matching what was actually running.
docker-compose.yml: added common_network to api-gateway.
These files now come from the call-center image (git-tracked as of
digiops/call-center 13ba8f8) instead of a host override that bypassed
CI. The same mounts also existed uncommitted in docker-compose.override.yml
on this host -- removed there too (not part of this commit, see local
working tree).