feat(gecx): stream BYOVA audio chunks

[adamweeks]

Summary

• stream each CES output audio frame immediately as raw 8 kHz mu-law BYOVA CHUNK audio
• keep START_OF_INPUT as a standalone event, then emit END_OF_INPUT, ordered audio chunks, and exactly one normal or terminal FINAL
• route the initial greeting through the same streaming path and remove GECX WAV buffering/playback delays
• reject unsupported GECX output formats until the later audio-format phase
• strengthen the GECX E2E plan so streamed-prompt pauses cannot produce false multi-turn passes

Validation

• 109 focused GECX/gateway tests
• 388 complete gateway tests
• 53 E2E framework tests
• four browser tests and production client build
• seven-scenario GECX plan validates with SHA-256 987a248a66aae1aaeb8b26dc58c05965cfd34d4df7e88ff01e8c2b67e556c127

Dev evidence

• corrected runtime commit 1611593 is active on the AWS dev instance; checksum, protobuf overlay/import preflight, service/public health, ListVirtualAgents, and both ALB targets passed
• long-response evidence emitted its first caller-turn audio chunk 1.016 seconds before CES marked the 31-chunk turn complete
• natural-pause evidence contains one intact caller turn in CES after a held speech end and merged resume
• strengthened multi-turn evidence contains both caller turns and both agent replies in CES
• task completion disconnected remotely with Webex code 0 and CES labels sessionContained=true/sessionEscalated=false
• transfer produced both announcements and CES labels sessionContained=false/sessionEscalated=true
• the timestamp-bounded corrected-release log scan contains no error, timeout, traceback, stream-error, or worker-join signature

The playback-speech scenario remains an audio-plane probe because barge-in is intentionally disabled until Phase 4.

Related: SE-1943