M4: voice pipeline (software-only telephony)

- AudioSocket server (asyncio TCP, Asterisk wire protocol) bridging calls
  into the same M2 agent used by chat
- Call session engine: greeting → energy-VAD utterance collection → STT →
  agent turn → TTS playback with barge-in (120ms of caller speech stops
  playback); inactivity + max-duration guards; unintelligible audio reaches
  the agent as '[nerazumljivo]' so the two-attempt rule stays in the prompt
- STTProvider (faster-whisper, lang hint 'sr', GPU/CPU via env) and
  TTSProvider (Azure Neural raw-8k PCM / ElevenLabs Flash) + test fakes
- Recording (mixed caller+agent WAV), transcripts, per-turn latency trace,
  metering via record_agent_call (§12)
- Internal dialplan API: /internal/calls/register (ring targets computed from
  working hours + ring settings §5.2), /answered (human outcome, no minutes),
  /hangup (abandoned → missed-call SMS §5.5); shared-token auth
- Asterisk config generator (pjsip.conf + extensions.conf from DB): worker
  endpoints, per-tenant test-caller endpoint, register→Dial→AudioSocket
  dialplan with h-extension reporting — validated against a real Asterisk 20
  container (modules, dialplan, endpoints all load)
- docker-compose 'voice' profile (voice server + Asterisk), Dockerfile.voice,
  docs/VOICE_TESTING.md softphone runbook
- 16 new tests incl. full fake-call e2e and a real-TCP AudioSocket wire test

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

gogo/voice/vad.py

@@ -0,0 +1,83 @@
+"""Utterance detection: adaptive energy VAD with hangover.
+
+Deliberately simple and dependency-free; swap for silero/webrtcvad later if the
+Phase-0 PoC shows it is needed. Operates on 20 ms SLIN frames.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from gogo.voice.audio import FRAME_MS, rms
+
+
+@dataclass
+class VadConfig:
+    # energy threshold = max(min_threshold, noise_floor * ratio)
+    min_threshold: int = 500
+    noise_ratio: float = 3.0
+    start_ms: int = 120  # this much voiced audio starts an utterance
+    end_ms: int = 700  # this much silence ends it
+    max_utterance_ms: int = 15000  # hard cap
+    pre_roll_ms: int = 200  # audio kept from before the trigger
+
+
+@dataclass
+class UtteranceCollector:
+    """Feed 20 ms frames; returns the full utterance PCM when one completes."""
+
+    config: VadConfig = field(default_factory=VadConfig)
+    _noise_floor: float = 200.0
+    _voiced_ms: int = 0
+    _silence_ms: int = 0
+    _in_utterance: bool = False
+    _buffer: bytearray = field(default_factory=bytearray)
+    _pre_roll: bytearray = field(default_factory=bytearray)
+
+    def is_speech(self, frame: bytes) -> bool:
+        energy = rms(frame)
+        threshold = max(self.config.min_threshold, self._noise_floor * self.config.noise_ratio)
+        speech = energy > threshold
+        if not speech:
+            # slowly track the noise floor on non-speech frames
+            self._noise_floor = 0.95 * self._noise_floor + 0.05 * energy
+        return speech
+
+    def feed(self, frame: bytes) -> bytes | None:
+        """Returns utterance PCM when a complete utterance is detected, else None."""
+        speech = self.is_speech(frame)
+
+        if not self._in_utterance:
+            self._pre_roll.extend(frame)
+            max_pre = self.config.pre_roll_ms * len(frame) // FRAME_MS
+            if len(self._pre_roll) > max_pre:
+                del self._pre_roll[: len(self._pre_roll) - max_pre]
+            if speech:
+                self._voiced_ms += FRAME_MS
+                if self._voiced_ms >= self.config.start_ms:
+                    self._in_utterance = True
+                    self._buffer = bytearray(self._pre_roll)
+                    self._silence_ms = 0
+            else:
+                self._voiced_ms = 0
+            return None
+
+        self._buffer.extend(frame)
+        if speech:
+            self._silence_ms = 0
+        else:
+            self._silence_ms += FRAME_MS
+
+        utterance_ms = len(self._buffer) * FRAME_MS // len(frame)
+        if self._silence_ms >= self.config.end_ms or utterance_ms >= self.config.max_utterance_ms:
+            utterance = bytes(self._buffer)
+            self.reset()
+            return utterance
+        return None
+
+    def reset(self) -> None:
+        self._in_utterance = False
+        self._voiced_ms = 0
+        self._silence_ms = 0
+        self._buffer = bytearray()
+        self._pre_roll = bytearray()