ci: use gpt-5.4-mini and hang up live calls

[dimavrem22]

Summary

• use gpt-5.4-mini for non-Realtime live intelligence while leaving Realtime media unchanged
• hang up every expected inbound and callback voice call through calls.hangup(call_id) in finally
• snapshot calls owned by each live pytest process, watchdog accidental calls in non-voice suites, and verify every new call is terminal during teardown
• require Inkbox SDK 0.4.24 for the external call-control API
• restore the inbound PSTN leg to required full-stack CI
• retain the mini-model prompt and behavioral assertion adjustments

Root cause fixed

The voice test finished before the driver's delayed media stop frame, and teardown killed the driver. The carrier leg then remained alive until the ten-minute cap. Call records also proved that general channel jobs can unexpectedly choose the phone tool, so cleanup limited to test_voice.py was insufficient.

Every live pytest process now owns a baseline of calls. Non-voice suites stop any newly created call within the watchdog interval; voice tests retain direct per-call cleanup and use the session guard as a fallback.

Validation

• live voice tests compile and collect successfully
• simulated cleanup preserved baseline/other-identity calls and terminated the newly owned call
• workflow YAML, prompt/tool-focused checks, and git diff --check passed
• full local suite: 283 passed; one unrelated QR-render test failed because the local environment lacks the declared Segno runtime