feat(voice): polish voice loop UX — filler kill, barge-in, DTX flush, time/RO TTS

End-to-end voice UX iteration after DAVE E2E shipped. Each change addresses a
real symptom Marius hit in live testing today:

- Kill the 3s filler ("mă gândesc"): Claude p50 is 4-7s so the filler always
  fired BEFORE the response and collided with it. Removed all filler infra
  from pipeline.py + tts_stream.py (FILLER_DELAY_S, _filler_task, push_filler,
  load_thinking_wav, thinking.wav cache).

- Barge-in: ttsq.clear() at the top of on_segment_done drops stale frames so
  a new utterance cuts off Echo's previous response cleanly.

- DTX silence flush: Discord stops sending RTP packets when the user goes
  silent (DTX), so the inline silence-check in sink.write() never fired for
  the trailing audio of an utterance — STT was missed entirely. Added a
  background poller thread that checks the silence-flush condition every
  200ms independent of incoming packets.

- Discord audio cadence fix: EchoStreamingAudioSource.read() blocked 100ms
  per call when pcm_queue was empty, wrecking Discord's 20ms frame pacing →
  client interpreted the stream as stutter and discarded leading frames
  (Marius heard "4 de minute în București" instead of the full sentence).
  Switched to get_frame_nowait() — instant return, silence frame on empty.

- RO time expansion: "23:09" was being read as "douăzeci și trei:nouă"
  with literal colon. Added expand_time() with feminine-correct minute
  formatting (un minut / două minute / douăzeci de minute / una de minute).

- Supertonic Unicode sanitize centralized in tools/tts.py: Romanian curly
  quotes (`„`, `"`, `"`, `—`, `…`) crash Supertonic with HTTP 500. Map them
  to ASCII at the synthesize() entry so BOTH voice mode and /audio command
  are covered without duplication. normalize.py re-exports for compat.

- Whisper offline: WhisperModel(..., local_files_only=True) — no more
  huggingface.co metadata GET on every startup. Model is already cached.

- Diagnostic logging across the chain: sink first-packet, VAD first-speech,
  voice stream block (Claude → callback), push_text (text → clauses queued),
  TTS pushed (clauses → frames). Lets future "spoke but Echo silent" bugs
  pinpoint exactly where the chain breaks.

- Captured Supertonic curly-quote lesson in tasks/lessons.md.

All 76 voice tests pass.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

cli.py

@@ -114,6 +114,104 @@ def _load_sessions_file() -> dict:
         return {}
 
 
+def _voice_doctor_checks() -> list[tuple[str, bool]]:
+    """Voice-stack health checks (Pas 10).
+
+    Mirrors the logic in tools/voice_setup.py but returns (label, ok) tuples
+    so they integrate with cmd_doctor's PASS/FAIL output. All checks degrade
+    gracefully — ImportError on optional voice deps is reported as FAIL, never
+    raised, so the rest of `eco doctor` is unaffected.
+    """
+    import importlib.util
+    import json as _json
+    import urllib.error
+    import urllib.request
+
+    results: list[tuple[str, bool]] = []
+
+    # 1. libopus0 loaded by discord.py
+    try:
+        import discord
+        if not discord.opus.is_loaded():
+            try:
+                discord.opus._load_default()
+            except Exception:
+                pass
+        results.append(("libopus loaded (discord.py)", discord.opus.is_loaded()))
+    except ImportError:
+        results.append(("libopus loaded (discord.py)", False))
+    except Exception:
+        results.append(("libopus loaded (discord.py)", False))
+
+    # 2. ffmpeg in PATH
+    results.append(("ffmpeg in PATH", shutil.which("ffmpeg") is not None))
+
+    # 3. Supertonic TTS reachable at http://127.0.0.1:7788/
+    supertonic_url = "http://127.0.0.1:7788/v1/audio/speech"
+    supertonic_ok = False
+    try:
+        payload = _json.dumps({
+            "model": "supertonic-3",
+            "input": "test",
+            "voice": "M2",
+            "response_format": "wav",
+            "lang": "ro",
+        }).encode("utf-8")
+        req = urllib.request.Request(
+            supertonic_url,
+            data=payload,
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        with urllib.request.urlopen(req, timeout=5) as resp:
+            supertonic_ok = resp.status == 200
+    except (urllib.error.URLError, ConnectionError, OSError):
+        supertonic_ok = False
+    except Exception:
+        supertonic_ok = False
+    results.append(("Supertonic TTS reachable at :7788", supertonic_ok))
+
+    # 4. faster-whisper importable (don't load model — too slow)
+    results.append((
+        "faster-whisper importable",
+        importlib.util.find_spec("faster_whisper") is not None,
+    ))
+
+    # 5. silero-vad importable
+    results.append((
+        "silero-vad importable",
+        importlib.util.find_spec("silero_vad") is not None,
+    ))
+
+    # 6. discord.ext.voice_recv importable (vendor package)
+    voice_recv_ok = False
+    try:
+        voice_recv_ok = importlib.util.find_spec("discord.ext.voice_recv") is not None
+    except (ImportError, ValueError, ModuleNotFoundError):
+        voice_recv_ok = False
+    except Exception:
+        voice_recv_ok = False
+    results.append(("discord.ext.voice_recv importable", voice_recv_ok))
+
+    # 7-9. Voice assets present and non-trivial size
+    voice_assets = [
+        ("assets/voice/thinking.wav", 1024),
+        ("assets/voice/beep_200ms.wav", 512),
+        ("assets/voice/mhm.wav", 512),
+    ]
+    for rel_path, min_bytes in voice_assets:
+        path = PROJECT_ROOT / rel_path
+        ok = False
+        try:
+            ok = path.exists() and path.stat().st_size > min_bytes
+        except OSError:
+            ok = False
+        label = f"{rel_path} (>{min_bytes}B)"
+        results.append((label, ok))
+
+    return results
+
+
 def cmd_doctor(args):
     """Run diagnostic checks."""
     import re
@@ -227,6 +325,9 @@ def cmd_doctor(args):
     else:
         checks.append(("WhatsApp bridge (optional)", True))
 
+    # ---- Voice stack checks (Pas 10) ----
+    checks.extend(_voice_doctor_checks())
+
     # Print results
     all_pass = True
     for label, passed in checks:

config.json

@@ -104,6 +104,12 @@
   "ollama": {
     "url": "http://10.0.20.161:11434"
   },
+  "voice": {
+    "allowed_user_ids": ["949388626146517022"],
+    "user_name": "Marius",
+    "default_voice": "M2",
+    "auto_leave_minutes": 5
+  },
   "paths": {
     "personality": "personality/",
     "tools": "tools/",

personality/AGENTS.md

@@ -189,4 +189,16 @@ Când lansez sub-agent, îi dau context: AGENTS.md, SOUL.md, USER.md + relevant
 - Discord links: `<url>` pentru a suprima embed-uri
 - Cand primesc o sarcina mai mare de executat, raspund intotdeauna cu o reactie sau confirmare si apoi trec la executie
 - **Link-uri:** Folosesc `https://moltbot.tailf7372d.ts.net/echo/` (NU IP 100.120.119.70) pentru ca WhatsApp să le recunoască ca link-uri
-- **Link-uri fișiere salvate:** Când salvez/menționez fișiere din `memory/kb/`, ofer automat link către `files.html#memory/kb/path/to/file.md` pentru preview
+- **Link-uri fișiere salvate:** Când salvez/menționez fișiere din `memory/kb/`, ofer automat link către `files.html#memory/kb/path/to/file.md` pentru preview
+
+## Voice mode
+
+Reguli aplicate când `adapter_name == "discord-voice"` — Marius mă ascultă, nu citește. Vocea e intolerantă la lung și la structură.
+
+- **1-3 propoziții max per răspuns.** Dacă am mai mult de spus, condensez sau mut în chat.
+- **Fără markdown.** Niciun bold, italic, cod cu backticks, headere. Text plat, atât.
+- **Fără bullet lists, nici numerotate.** Le pronunț natural ca propoziții: "trei lucruri: în primul rând..., apoi..., și la final..."
+- **Fără linkuri.** Nu rostesc URL-uri. Dacă e relevant: "îți trimit linkul în chat".
+- **Numere și valute formulate conversațional.** Scriu "treizeci de lei", nu "30 RON"; "douăzeci și cinci la sută", nu "25%". Modulul `normalize.py` face curățare tehnică, dar eu formulez deja natural — un om vorbește, nu citește tabelul.
+- **Lung sau structurat → mută în chat.** Dacă răspunsul cere listă, cod, linkuri sau peste 3 propoziții, închei rostit cu "L-am scris în chat." iar restul ajunge în text channel mirror.
+- **Ton:** cum vorbesc cu Marius la o cafea, nu cum scriu raport. Contracții, pauze, "păi" sau "stai puțin" dacă mă ajută să sune uman. Concis, fără tic-uri robotice.

personality/TOOLS.md

@@ -63,6 +63,13 @@
 - **Venv:** ~/echo-core/.venv/ | **Model:** base
 - **Utilizare:** `whisper.load_model('base').transcribe(path, language='ro')`
 
+### Discord Voice
+- **Ce este:** Bot conectat la un voice channel Discord — ascultă microfonul lui Marius, transcrie cu faster-whisper (`small` int8, RO), rutează prin router și răspunde rostit cu Supertonic TTS.
+- **Cum sunt "în voce":** Slash command `/voice join` mă cheamă în channel; cât stau acolo, presence-ul arată că ascult. `/voice leave` sau auto-leave după 5 minute fără voce.
+- **Latență așteptată:** ~5 secunde perceput end-to-end (STT p50 2.25s + LLM + TTS first chunk). Peste 3s pornesc un filler audio ("Stai să-mi adun gândurile") ca să nu pară mort.
+- **Streaming TTS:** răspunsul iese pe clauze, nu cuvânt-cu-cuvânt și nu frază întreagă — primul sunet pleacă imediat ce am o propoziție scurtă.
+- **Limitări:** 1-3 propoziții max (vezi AGENTS.md § Voice mode). Cuvinte rare, nume proprii sau acronime pot apărea ciudat în STT — dacă sună greșit, cer reformulare în loc să ghicesc.
+
 ### Pauze respirație
 - **Script:** `python3 tools/pauza_random.py`
 - **Bancă:** memory/kb/tehnici-pauza.md

src/adapters/discord_bot.py

@@ -112,6 +112,7 @@ def create_bot(config: Config) -> discord.Client:
 
     intents = discord.Intents.default()
     intents.message_content = True
+    intents.voice_states = True
 
     client = discord.Client(intents=intents)
     tree = app_commands.CommandTree(client)
@@ -958,6 +959,11 @@ def create_bot(config: Config) -> discord.Client:
         else:
             await interaction.followup.send(result or "Eroare TTS.")
 
+    # Voice slash group (Pas 7)
+    from src.adapters.discord_voice import register as register_voice
+    voice_group = register_voice(tree, client)
+    tree.add_command(voice_group)
+
     # --- Ralph commands (autonomous project execution) ---
 
     async def _autocomplete_by_status(
@@ -1118,6 +1124,11 @@ def create_bot(config: Config) -> discord.Client:
         from datetime import datetime, timezone
         client._ready_at = datetime.now(timezone.utc)
         logger.info("Echo Core online as %s", client.user)
+        # Voice models eager warmup (Pas 7)
+        from src.adapters import discord_voice
+        discord_voice._models_warmup_future = asyncio.create_task(
+            discord_voice.warmup_models()
+        )
 
     async def _handle_chat(message: discord.Message) -> None:
         """Process a chat message through the router and send the response."""

src/adapters/discord_voice.py

@@ -0,0 +1,322 @@
+"""Discord voice slash commands (Pas 7 — CONVERGENCE wiring).
+
+Registers the `/voice` slash command group on the existing CommandTree and
+exposes an async `warmup_models()` for eager model load at bot startup.
+
+Owns nothing in `src/voice/*` — purely the Discord-facing wiring. Defers
+heavy lifting to:
+
+- ``src.voice.pipeline.VoiceSession`` — per-guild session state machine
+- ``src.voice.pipeline.EchoVoiceSink`` — discord-ext-voice-recv sink
+- ``src.voice.tts_stream.TTSQueue`` / ``EchoStreamingAudioSource``
+- ``src.voice._discord_voice_adapter.connect_voice``
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import Optional
+
+import discord
+from discord import app_commands
+
+# Optional DAVE dep (mandatory at runtime when discord.py 2.7.1 is paired with
+# Discord voice gateway v=8; tolerated missing in tests / dev environments).
+try:
+    import davey
+    _HAS_DAVE = True
+except ImportError:
+    _HAS_DAVE = False
+
+from src.config import Config
+from src.voice.pipeline import (
+    VoiceSession,
+    EchoVoiceSink,
+    _get_whisper_model,
+    _get_silero_vad,
+)
+from src.voice.tts_stream import TTSQueue, EchoStreamingAudioSource
+from src.voice._discord_voice_adapter import connect_voice
+
+log = logging.getLogger("echo-core.discord.voice")
+
+# Per-guild voice session registry. Key = guild_id.
+_voice_sessions: dict[int, VoiceSession] = {}
+
+# Set if model warmup failed; surfaces as ephemeral error on /voice join.
+_voice_load_error: Optional[str] = None
+
+# Reference to the eager warmup task created in on_ready, so /voice join can
+# await it if the user is faster than the background load.
+_models_warmup_future: Optional[asyncio.Task] = None
+
+
+async def warmup_models() -> None:
+    """Eager model load — called from `on_ready()` as a background task.
+
+    Runs the (synchronous, blocking) model loaders on a worker thread so the
+    event loop stays responsive. On failure, sets `_voice_load_error` instead
+    of raising, so `/voice join` can degrade gracefully.
+    """
+    global _voice_load_error
+    try:
+        if not discord.opus.is_loaded():
+            discord.opus.load_opus("libopus.so.0")
+        if _HAS_DAVE:
+            log.info("DAVE protocol v%d available (davey %s)",
+                     davey.DAVE_PROTOCOL_VERSION, davey.__version__)
+        await asyncio.to_thread(_get_whisper_model)
+        await asyncio.to_thread(_get_silero_vad)
+        log.info("Voice models warm")
+    except Exception as e:
+        _voice_load_error = f"{type(e).__name__}: {e}"
+        log.error("Voice models load failed: %s", _voice_load_error)
+
+
+def _get_whitelist() -> set[int]:
+    """Read `voice.allowed_user_ids` from config and coerce to int set.
+
+    Re-reads config from disk to pick up any runtime edits between bot start
+    and /voice join.
+    """
+    try:
+        raw = Config().get("voice.allowed_user_ids", [])
+    except Exception:
+        raw = []
+    out: set[int] = set()
+    for v in raw or []:
+        try:
+            out.add(int(v))
+        except (TypeError, ValueError):
+            continue
+    return out
+
+
+def _get_default_voice() -> str:
+    try:
+        return Config().get("voice.default_voice", "M2") or "M2"
+    except Exception:
+        return "M2"
+
+
+def register(tree: app_commands.CommandTree, bot: discord.Client) -> app_commands.Group:
+    """Build the `/voice` slash command group and return it (caller registers)."""
+    voice_group = app_commands.Group(
+        name="voice", description="Echo Core voice channel"
+    )
+
+    @voice_group.command(name="join", description="Echo intră în voice channel-ul tău")
+    async def join(interaction: discord.Interaction) -> None:
+        await interaction.response.defer(ephemeral=True)
+        if _voice_load_error:
+            await interaction.followup.send(
+                f"Voice unavailable: {_voice_load_error}", ephemeral=True
+            )
+            return
+        if _models_warmup_future is not None and not _models_warmup_future.done():
+            try:
+                await _models_warmup_future
+            except Exception as e:
+                await interaction.followup.send(
+                    f"Voice unavailable: {type(e).__name__}: {e}", ephemeral=True
+                )
+                return
+        user = interaction.user
+        if not isinstance(user, discord.Member) or user.voice is None or user.voice.channel is None:
+            await interaction.followup.send(
+                "Intră într-un voice channel întâi.", ephemeral=True
+            )
+            return
+        channel = user.voice.channel
+        whitelist = _get_whitelist()
+        if user.id not in whitelist:
+            await interaction.followup.send(
+                "Nu ești pe whitelist voice.", ephemeral=True
+            )
+            return
+        # Reject double-join on the same guild.
+        guild_id = channel.guild.id
+        if guild_id in _voice_sessions:
+            await interaction.followup.send(
+                "Sunt deja în voice pe acest server. Folosește /voice leave întâi.",
+                ephemeral=True,
+            )
+            return
+        # Connect
+        try:
+            vc = await connect_voice(channel)
+        except Exception as e:
+            log.exception("connect_voice failed")
+            await interaction.followup.send(
+                f"Conectare eșuată: {type(e).__name__}: {e}", ephemeral=True
+            )
+            return
+        # Build TTS queue + session
+        ttsq = TTSQueue(voice_id=_get_default_voice(), lang="ro")
+        ttsq.start()
+        try:
+            session = VoiceSession(
+                channel_id=channel.id,
+                guild_id=guild_id,
+                voice_client=vc,
+                text_channel=interaction.channel,
+                record_enabled=False,
+                mirror_enabled=True,
+                whitelist=whitelist,
+                ttsq=ttsq,
+                bot=bot,
+                loop=asyncio.get_running_loop(),
+            )
+        except Exception as e:
+            log.exception("VoiceSession construction failed")
+            ttsq.stop()
+            try:
+                await vc.disconnect(force=True)
+            except Exception:
+                pass
+            await interaction.followup.send(
+                f"Sesiune voice eșuată: {type(e).__name__}: {e}", ephemeral=True
+            )
+            return
+        _voice_sessions[guild_id] = session
+        # Start TTS streaming source for the entire session. Chain the
+        # wake-up beep via `after=` so streaming takes over when beep ends.
+        def _start_stream(error: Optional[Exception] = None) -> None:
+            if error is not None:
+                log.warning("Beep playback ended with error: %s", error)
+            try:
+                vc.play(EchoStreamingAudioSource(ttsq))
+                log.info("TTS streaming source attached")
+            except Exception:
+                log.exception("EchoStreamingAudioSource attach failed")
+        try:
+            vc.play(
+                discord.FFmpegPCMAudio("assets/voice/beep_200ms.wav"),
+                after=_start_stream,
+            )
+        except Exception:
+            log.warning("Beep playback skipped, starting stream directly", exc_info=True)
+            _start_stream()
+        # Attach sink
+        try:
+            bot_user_id = int(bot.user.id) if bot.user is not None else 0
+            sink = EchoVoiceSink(session=session, bot_user_id=bot_user_id)
+            vc.listen(sink)
+        except Exception as e:
+            log.exception("Sink attach failed")
+            _voice_sessions.pop(guild_id, None)
+            try:
+                session.cleanup("sink_attach_failed")
+            except Exception:
+                pass
+            await interaction.followup.send(
+                f"Atașare sink eșuată: {type(e).__name__}: {e}", ephemeral=True
+            )
+            return
+        # Presence
+        try:
+            await bot.change_presence(activity=discord.Activity(
+                type=discord.ActivityType.listening,
+                name=f"{user.display_name} în #{channel.name}",
+            ))
+        except Exception:
+            log.warning("Presence update skipped", exc_info=True)
+        await interaction.followup.send(
+            f"În voce în #{channel.name}.", ephemeral=True
+        )
+
+    @voice_group.command(name="leave", description="Echo iese din voice channel")
+    async def leave(interaction: discord.Interaction) -> None:
+        await interaction.response.defer(ephemeral=True)
+        guild_id = interaction.guild.id if interaction.guild else None
+        session = _voice_sessions.pop(guild_id, None) if guild_id is not None else None
+        if session is None:
+            await interaction.followup.send(
+                "Nu sunt în niciun voice channel aici.", ephemeral=True
+            )
+            return
+        try:
+            session.cleanup("user_leave")
+        except Exception:
+            log.exception("session.cleanup raised")
+        try:
+            await bot.change_presence(activity=None)
+        except Exception:
+            log.warning("Presence reset skipped", exc_info=True)
+        await interaction.followup.send("Plecat.", ephemeral=True)
+
+    @voice_group.command(name="doctor", description="Verifică voice stack")
+    async def doctor(interaction: discord.Interaction) -> None:
+        await interaction.response.defer(ephemeral=True)
+        checks: list[tuple[str, bool]] = []
+        # libopus
+        try:
+            checks.append(("libopus", bool(discord.opus.is_loaded())))
+        except Exception:
+            checks.append(("libopus", False))
+        # warmup
+        checks.append(("voice load error", _voice_load_error is None))
+        # Build response
+        lines = ["**Voice doctor:**"]
+        for label, ok in checks:
+            lines.append(f"{'OK' if ok else 'FAIL'} — {label}")
+        if _voice_load_error:
+            lines.append(f"  details: {_voice_load_error}")
+        await interaction.followup.send("\n".join(lines), ephemeral=True)
+
+    # --- /voice mirror on|off ---
+    mirror_group = app_commands.Group(
+        name="mirror", description="Text mirror", parent=voice_group
+    )
+
+    @mirror_group.command(name="on", description="Activează text mirror în canal")
+    async def mirror_on(interaction: discord.Interaction) -> None:
+        await interaction.response.defer(ephemeral=True)
+        guild_id = interaction.guild.id if interaction.guild else None
+        s = _voice_sessions.get(guild_id) if guild_id is not None else None
+        if s is None:
+            await interaction.followup.send("Nu sunt în voice.", ephemeral=True)
+            return
+        s.mirror_enabled = True
+        await interaction.followup.send("Mirror ON.", ephemeral=True)
+
+    @mirror_group.command(name="off", description="Dezactivează text mirror")
+    async def mirror_off(interaction: discord.Interaction) -> None:
+        await interaction.response.defer(ephemeral=True)
+        guild_id = interaction.guild.id if interaction.guild else None
+        s = _voice_sessions.get(guild_id) if guild_id is not None else None
+        if s is None:
+            await interaction.followup.send("Nu sunt în voice.", ephemeral=True)
+            return
+        s.mirror_enabled = False
+        await interaction.followup.send("Mirror OFF.", ephemeral=True)
+
+    # --- /voice record on|off ---
+    record_group = app_commands.Group(
+        name="record", description="KB recording", parent=voice_group
+    )
+
+    @record_group.command(name="on", description="Activează înregistrare în KB")
+    async def record_on(interaction: discord.Interaction) -> None:
+        await interaction.response.defer(ephemeral=True)
+        guild_id = interaction.guild.id if interaction.guild else None
+        s = _voice_sessions.get(guild_id) if guild_id is not None else None
+        if s is None:
+            await interaction.followup.send("Nu sunt în voice.", ephemeral=True)
+            return
+        s.record_enabled = True
+        await interaction.followup.send("Record ON.", ephemeral=True)
+
+    @record_group.command(name="off", description="Dezactivează înregistrare")
+    async def record_off(interaction: discord.Interaction) -> None:
+        await interaction.response.defer(ephemeral=True)
+        guild_id = interaction.guild.id if interaction.guild else None
+        s = _voice_sessions.get(guild_id) if guild_id is not None else None
+        if s is None:
+            await interaction.followup.send("Nu sunt în voice.", ephemeral=True)
+            return
+        s.record_enabled = False
+        await interaction.followup.send("Record OFF.", ephemeral=True)
+
+    return voice_group

src/claude_session.py

@@ -37,6 +37,42 @@ DEFAULT_TIMEOUT = 300  # seconds
 
 CLAUDE_BIN = os.environ.get("CLAUDE_BIN", "claude")
 
+# ---------------------------------------------------------------------------
+# Per-channel mutex for send_message
+# ---------------------------------------------------------------------------
+#
+# Two paths can hit `send_message(channel_id, ...)` concurrently for the same
+# channel: a text adapter (Discord/Telegram/WhatsApp) and the voice adapter
+# (`adapter_name="discord-voice"`). The underlying Claude CLI subprocess is
+# blocking (`subprocess.Popen` with stream-json read loop) and stateful via
+# `--resume <session_id>` — interleaving two concurrent invocations on the
+# same channel would corrupt the conversation order.
+#
+# We use `threading.Lock` (NOT `asyncio.Lock`) because `send_message` is sync
+# code typically run from `asyncio.to_thread` in async adapters. asyncio.Lock
+# only serializes coroutines, not threads — it would NOT protect this path.
+#
+# Each channel gets its own lock so DIFFERENT channels still run in parallel.
+# Locks are created lazily on first use; the dict itself is guarded by a
+# small bootstrap lock so two concurrent first-uses don't race on creation.
+_session_locks: dict[str, threading.Lock] = {}
+_session_locks_bootstrap = threading.Lock()
+
+
+def _get_session_lock(channel_id: str) -> threading.Lock:
+    """Return the channel's mutex, creating it on first access.
+
+    Two threads racing to create the same channel's lock would otherwise
+    end up with different lock objects (setdefault is not atomic across
+    the read-modify-write under all interpreter conditions — defensive).
+    """
+    lock = _session_locks.get(channel_id)
+    if lock is not None:
+        return lock
+    with _session_locks_bootstrap:
+        return _session_locks.setdefault(channel_id, threading.Lock())
+
+
 PERSONALITY_FILES = [
     "IDENTITY.md",
     "SOUL.md",
@@ -543,19 +579,28 @@ def send_message(
     timeout: int = DEFAULT_TIMEOUT,
     on_text: Callable[[str], None] | None = None,
 ) -> str:
-    """High-level convenience: auto start or resume based on channel state."""
-    session = get_active_session(channel_id)
-    # Only resume if session has a valid session_id (not a pre-set model placeholder)
-    if session is not None and session.get("session_id"):
-        return resume_session(session["session_id"], message, timeout, on_text=on_text)
-    # Use model from pre-set session if available, otherwise use provided model
-    effective_model = model
-    if session is not None and session.get("model"):
-        effective_model = session["model"]
-    response_text, _session_id = start_session(
-        channel_id, message, effective_model, timeout, on_text=on_text
-    )
-    return response_text
+    """High-level convenience: auto start or resume based on channel state.
+
+    Concurrency: a per-`channel_id` `threading.Lock` serializes invocations
+    that hit the same channel (e.g. text adapter + voice adapter racing on
+    the same Discord guild text channel). Different channels run in
+    parallel — each holds its own lock. Lock is acquired blocking; we rely
+    on `timeout` (default 5 minutes) to bound the worst case rather than
+    a non-blocking acquire (loss of fairness vs adapter-side queueing).
+    """
+    with _get_session_lock(channel_id):
+        session = get_active_session(channel_id)
+        # Only resume if session has a valid session_id (not a pre-set model placeholder)
+        if session is not None and session.get("session_id"):
+            return resume_session(session["session_id"], message, timeout, on_text=on_text)
+        # Use model from pre-set session if available, otherwise use provided model
+        effective_model = model
+        if session is not None and session.get("model"):
+            effective_model = session["model"]
+        response_text, _session_id = start_session(
+            channel_id, message, effective_model, timeout, on_text=on_text
+        )
+        return response_text
 
 
 def clear_session(channel_id: str) -> bool:

src/router.py

@@ -154,8 +154,17 @@ def route_message(
             channel_cfg = _get_channel_config(channel_id)
             model = (channel_cfg or {}).get("default_model") or _get_config().get("bot.default_model", "sonnet")
 
+    # Voice-mode augment: prepend speaker prefix so Claude knows who spoke
+    # in a voice channel. Cheap now, future-proof for multi-speaker later.
+    # (Engineering decision #14 in the plan.) Only the discord-voice adapter
+    # triggers it — text adapters keep the message verbatim.
+    claude_text = text
+    if adapter_name == "discord-voice":
+        user_name = _get_config().get("voice.user_name", "user") or "user"
+        claude_text = f"[speaker:{user_name}] {text}"
+
     try:
-        response = send_message(channel_id, text, model=model, on_text=on_text)
+        response = send_message(channel_id, claude_text, model=model, on_text=on_text)
         _set_last_response(channel_id, response)
         return response, False
     except Exception as e:

src/voice/_discord_voice_adapter.py

@@ -0,0 +1,67 @@
+# -*- coding: utf-8 -*-
+"""Adapter layer over `discord-ext-voice-recv` (vendored at vendor/).
+
+If discord-ext-voice-recv breaks, swap to py-cord by rewriting only this file.
+Contract test in tests/test_voice_adapter_contract.py guards drift.
+
+Downstream consumers (`src/voice/*`, `src/adapters/discord_voice.py`) MUST
+import from this file — never from `discord.ext.voice_recv` directly.
+
+## Public API surface (stable across upstream changes)
+
+- ``VoiceReceiveClient`` — alias for ``voice_recv.VoiceRecvClient``. Subclass
+  of ``discord.VoiceClient`` with extra audio-receive plumbing.
+  Key methods used by the pipeline:
+    * ``await client.disconnect(force: bool = False)`` (from discord.VoiceClient)
+    * ``client.listen(sink, *, after=None)`` — attach an ``AudioSink``;
+      raises ``discord.ClientException`` if not connected or already listening
+    * ``client.stop_listening()`` — detach the current sink
+    * ``client.is_listening() -> bool``
+    * ``client.stop()`` — stop both playing and listening
+    * ``client.sink`` (property, getter+setter) — swap the active sink in place
+
+- ``AudioSink`` — abstract base. Subclasses MUST implement:
+    * ``write(user: Optional[discord.User|Member], data: VoiceData) -> None``
+    * ``wants_opus() -> bool`` (True → receive opus bytes; False → receive PCM)
+    * ``cleanup() -> None``
+
+- ``VoiceData`` — per-packet container. Slots: ``packet``, ``source``, ``pcm``.
+  ``.pcm`` is decoded 48kHz s16le stereo bytes when ``wants_opus()`` is False.
+  ``.opus`` property returns the raw opus bytes from the underlying RTP packet.
+
+- ``connect_voice(channel) -> VoiceReceiveClient`` — async helper, returns a
+  connected receive-capable voice client.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from discord.ext import voice_recv
+
+if TYPE_CHECKING:
+    import discord
+
+
+# --- Stable re-exports -------------------------------------------------------
+
+VoiceReceiveClient = voice_recv.VoiceRecvClient
+AudioSink = voice_recv.AudioSink
+VoiceData = voice_recv.VoiceData
+
+
+__all__ = [
+    "VoiceReceiveClient",
+    "AudioSink",
+    "VoiceData",
+    "connect_voice",
+]
+
+
+async def connect_voice(channel: "discord.VoiceChannel") -> VoiceReceiveClient:
+    """Connect to a Discord voice channel with the receive-capable client.
+
+    Thin wrapper around ``channel.connect(cls=VoiceRecvClient)`` so callers
+    don't have to import the vendored class directly.
+    """
+    return await channel.connect(cls=VoiceReceiveClient)

src/voice/normalize.py

@@ -0,0 +1,267 @@
+"""Voice mode text normalization for Romanian TTS.
+
+Pure functions — no side effects, no I/O, no logging. Strip markdown,
+expand numbers / currency / symbols / abbreviations into natural-sounding
+Romanian text. See plan: src/voice/normalize.py (Pas 3).
+
+Pipeline order in normalize_for_tts:
+    strip_markdown -> expand_abbreviations -> expand_currency
+    -> expand_numbers_ro -> expand_symbols -> truncate(200)
+
+Currency runs BEFORE generic number expansion so "12.50 RON" becomes
+"doisprezece lei și cincizeci de bani" rather than
+"doisprezece virgulă cincizeci RON".
+"""
+import re
+
+from num2words import num2words
+
+
+# ---------- Markdown ----------
+
+_MARKDOWN_LINK = re.compile(r'\[([^\]]+)\]\([^)]+\)')
+_MARKDOWN_BOLD = re.compile(r'\*\*([^*]+)\*\*')
+_MARKDOWN_CODE = re.compile(r'`([^`\n]+)`')
+_MARKDOWN_ITALIC = re.compile(r'(?<!\*)\*([^*\n]+)\*(?!\*)')
+_MARKDOWN_HEADING = re.compile(r'^[ \t]*#{1,6}[ \t]+', re.MULTILINE)
+_MARKDOWN_LIST = re.compile(r'^[ \t]*[-*+][ \t]+', re.MULTILINE)
+
+
+def strip_markdown(text: str) -> str:
+    """Remove common markdown formatting, preserve the visible content."""
+    text = _MARKDOWN_LINK.sub(r'\1', text)
+    text = _MARKDOWN_BOLD.sub(r'\1', text)
+    text = _MARKDOWN_CODE.sub(r'\1', text)
+    text = _MARKDOWN_ITALIC.sub(r'\1', text)
+    text = _MARKDOWN_HEADING.sub('', text)
+    text = _MARKDOWN_LIST.sub('', text)
+    return text
+
+
+# ---------- Number helpers ----------
+
+def _needs_de(n: int) -> bool:
+    """Romanian: insert 'de' between numeral and noun for n >= 20,
+    except when the trailing 1-19 portion makes it sound off
+    (e.g., 105, 119 -> no 'de'; 120, 200 -> 'de').
+    """
+    if n < 20:
+        return False
+    last = n % 100
+    if 1 <= last <= 19:
+        return False
+    return True
+
+
+def _int_to_ro(n: int) -> str:
+    return num2words(n, lang='ro')
+
+
+def _decimal_to_ro(s: str) -> str:
+    """Convert decimal string 'X.Y' to RO words.
+
+    Decimal part is read as a whole number ('3.14' -> 'trei virgulă paisprezece'),
+    unless it has a leading zero ('3.05' -> 'trei virgulă zero cinci') so the
+    magnitude is preserved.
+    """
+    int_part, dec_part = s.split('.', 1)
+    int_words = _int_to_ro(int(int_part))
+    if dec_part.startswith('0') and len(dec_part) > 1:
+        dec_words = ' '.join(_int_to_ro(int(d)) for d in dec_part)
+    else:
+        dec_words = _int_to_ro(int(dec_part))
+    return f"{int_words} virgulă {dec_words}"
+
+
+# ---------- Numbers ----------
+
+_NUM_TOKEN = re.compile(r'(?<!\w)(\d+(?:\.\d+)?)(?!\w)')
+
+
+def expand_numbers_ro(text: str) -> str:
+    """Expand bare numeric tokens to Romanian words.
+
+    Only matches pure number tokens (no surrounding letters). Decimals
+    use 'virgulă' separator. Currency-bound numbers should already be
+    handled by expand_currency before this runs.
+    """
+    def _sub(match: re.Match) -> str:
+        token = match.group(1)
+        if '.' in token:
+            return _decimal_to_ro(token)
+        return _int_to_ro(int(token))
+
+    return _NUM_TOKEN.sub(_sub, text)
+
+
+# ---------- Time ----------
+
+_TIME_PATTERN = re.compile(r'(?<!\d)([01]?\d|2[0-3]):([0-5]?\d)(?!\d)')
+
+
+def _format_minutes_ro(n: int) -> str:
+    """Romanian-correct feminine forms for minute counts (0-59)."""
+    if n == 1:
+        return "un minut"
+    if n == 2:
+        return "două minute"
+    if n < 20:
+        return f"{_int_to_ro(n)} minute"
+    last = n % 10
+    rest = n - last
+    if last == 0:
+        return f"{_int_to_ro(n)} de minute"
+    if last == 1:
+        return f"{_int_to_ro(rest)} și una de minute"
+    if last == 2:
+        return f"{_int_to_ro(rest)} și două de minute"
+    return f"{_int_to_ro(rest)} și {_int_to_ro(last)} de minute"
+
+
+def expand_time(text: str) -> str:
+    """Expand ``HH:MM`` clock times into colloquial Romanian.
+
+    23:09 -> "douăzeci și trei și nouă minute"
+    23:00 -> "douăzeci și trei fix"
+    """
+    def _sub(match: re.Match) -> str:
+        h = int(match.group(1))
+        m = int(match.group(2))
+        hour_str = _int_to_ro(h)
+        if m == 0:
+            return f"{hour_str} fix"
+        return f"{hour_str} și {_format_minutes_ro(m)}"
+
+    return _TIME_PATTERN.sub(_sub, text)
+
+
+# ---------- Currency ----------
+
+_CURRENCY_MAIN = {
+    'RON': ('leu', 'lei'),
+    'USD': ('dolar', 'dolari'),
+    'EUR': ('euro', 'euro'),
+    'GBP': ('liră', 'lire'),
+}
+
+_CURRENCY_SUB = {
+    'RON': ('ban', 'bani'),
+    'USD': ('cent', 'cenți'),
+    'EUR': ('cent', 'cenți'),
+    'GBP': ('penny', 'pence'),
+}
+
+_CURRENCY_PATTERNS = [
+    # RON suffix (case-insensitive: RON, ron, lei)
+    (re.compile(r'(?<!\w)(\d+(?:\.\d+)?)\s+(?:RON|lei)\b', re.IGNORECASE), 'RON'),
+    # Prefix currencies
+    (re.compile(r'\$(\d+(?:\.\d+)?)'), 'USD'),
+    (re.compile(r'€(\d+(?:\.\d+)?)'), 'EUR'),
+    (re.compile(r'£(\d+(?:\.\d+)?)'), 'GBP'),
+]
+
+
+def _format_currency_unit(n: int, singular: str, plural: str) -> str:
+    """Format integer amount + currency noun with proper RO singular/plural
+    and 'de' particle. Uses 'un' (article) for n=1, not 'unu' (cardinal).
+    """
+    if n == 1:
+        return f"un {singular}"
+    word = _int_to_ro(n)
+    if _needs_de(n):
+        return f"{word} de {plural}"
+    return f"{word} {plural}"
+
+
+def _format_currency(amount: str, code: str) -> str:
+    main_sg, main_pl = _CURRENCY_MAIN[code]
+    if '.' in amount:
+        whole_s, frac_s = amount.split('.', 1)
+        # Normalize fractional part to 2 digits so "12.5 RON" reads as
+        # 50 bani, not 5 bani.
+        if len(frac_s) == 1:
+            frac_s = frac_s + '0'
+        elif len(frac_s) > 2:
+            frac_s = frac_s[:2]
+        whole = int(whole_s)
+        frac = int(frac_s)
+        whole_part = _format_currency_unit(whole, main_sg, main_pl)
+        if frac == 0:
+            return whole_part
+        sub_sg, sub_pl = _CURRENCY_SUB[code]
+        frac_part = _format_currency_unit(frac, sub_sg, sub_pl)
+        return f"{whole_part} și {frac_part}"
+    return _format_currency_unit(int(amount), main_sg, main_pl)
+
+
+def expand_currency(text: str) -> str:
+    """Expand currency amounts into natural Romanian.
+
+    Recognises ``<n> RON`` / ``<n> lei`` suffix and ``$``, ``€``, ``£`` prefix
+    forms with optional 2-decimal fractional part (treated as sub-unit:
+    bani / cenți / pence).
+    """
+    for pattern, code in _CURRENCY_PATTERNS:
+        text = pattern.sub(lambda m, c=code: _format_currency(m.group(1), c), text)
+    return text
+
+
+# ---------- Symbols ----------
+
+def expand_symbols(text: str) -> str:
+    """Replace common symbols with their Romanian spoken form."""
+    text = text.replace('%', ' la sută')
+    text = text.replace('&', ' și ')
+    text = text.replace('@', ' la ')
+    text = text.replace('°', ' grade')
+    text = re.sub(r'\s+', ' ', text).strip()
+    return text
+
+
+from tools.tts import sanitize_for_supertonic as sanitize_punctuation
+
+
+# ---------- Abbreviations ----------
+
+# Longer patterns first so 'ș.a.m.d.' wins over 'ș.a.'
+_ABBREVIATIONS = [
+    (re.compile(r'(?<!\w)[șş]\.a\.m\.d\.', re.IGNORECASE), 'și așa mai departe'),
+    (re.compile(r'(?<!\w)[șş]\.a\.', re.IGNORECASE), 'și altele'),
+    (re.compile(r'(?<!\w)etc\.', re.IGNORECASE), 'etcetera'),
+    (re.compile(r'(?<!\w)dl\.', re.IGNORECASE), 'domnul'),
+    (re.compile(r'(?<!\w)dna\.', re.IGNORECASE), 'doamna'),
+    (re.compile(r'(?<!\w)nr\.', re.IGNORECASE), 'numărul'),
+]
+
+
+def expand_abbreviations(text: str) -> str:
+    """Expand Romanian abbreviations into their full forms."""
+    for pattern, replacement in _ABBREVIATIONS:
+        text = pattern.sub(replacement, text)
+    return text
+
+
+# ---------- Top-level pipeline ----------
+
+_MAX_WORDS = 200
+_TRUNCATE_SUFFIX = "Restul l-am scris în chat."
+
+
+def normalize_for_tts(text: str) -> str:
+    """Apply the full normalization pipeline and truncate to 200 words.
+
+    If the text exceeds 200 words, the first 200 are kept and the suffix
+    "Restul l-am scris în chat." is appended so the listener knows the
+    response continues in the text channel mirror.
+    """
+    text = strip_markdown(text)
+    text = sanitize_punctuation(text)
+    text = expand_abbreviations(text)
+    text = expand_time(text)
+    text = expand_currency(text)
+    text = expand_numbers_ro(text)
+    text = expand_symbols(text)
+    words = text.split()
+    if len(words) > _MAX_WORDS:
+        text = ' '.join(words[:_MAX_WORDS]) + f" {_TRUNCATE_SUFFIX}"
+    return text.strip()

src/voice/pipeline.py

@@ -0,0 +1,580 @@
+"""Central voice pipeline: VAD -> STT -> Claude -> TTS for Discord voice.
+
+``VoiceSession`` binds per-call state — voice_client, TTS queue, transcript
+JSONL buffer, whitelist, presence — and exposes a single idempotent
+``cleanup()`` invoked from every exit path (user /voice leave, network
+disconnect, crash via ``__exit__``, auto-leave timer, user leaves channel).
+
+``EchoVoiceSink`` is the discord-ext-voice-recv ``AudioSink`` subclass that
+runs in the voice_recv reader thread. It batches 20ms PCM packets into
+100ms windows for silero-vad inference, marks per-user speech timestamps,
+and on 800ms cumulative silence flushes the accumulated audio through
+faster-whisper. Hallucinated segments (``no_speech_prob > 0.6``) are
+dropped. Valid transcripts are scheduled onto the session's event loop
+via ``asyncio.run_coroutine_threadsafe``.
+
+The bot's own ``user.id`` is filtered FIRST inside ``write()`` — load-bearing
+echo prevention so a future whitelist expansion (Bianca, etc.) never lets
+the bot transcribe itself.
+
+See plan: ``src/voice/pipeline.py`` (Pas 5), Engineering decisions #4
+(VAD 100ms batched), #5 (cleanup centralizat), #7 (bot.user.id explicit
+guard).
+"""
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import threading
+import time
+from pathlib import Path
+from typing import Any, Callable, Optional
+
+import numpy as np
+
+from src.voice._discord_voice_adapter import AudioSink, VoiceData
+
+log = logging.getLogger(__name__)
+
+
+# Discord delivers 48kHz s16le stereo PCM, 20ms per packet (3840 bytes).
+SAMPLE_RATE_DISCORD = 48000
+SAMPLE_RATE_WHISPER = 16000
+PACKET_MS = 20
+PACKET_BYTES = 3840  # 48000 Hz * 0.020 s * 2 channels * 2 bytes
+VAD_WINDOW_MS = 100  # batch 5 * 20ms packets per VAD inference (Decision #4)
+VAD_WINDOW_BYTES = PACKET_BYTES * (VAD_WINDOW_MS // PACKET_MS)
+VAD_THRESHOLD = 0.5
+SILENCE_FLUSH_MS = 800
+NO_SPEECH_DROP_THRESHOLD = 0.6
+
+PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent
+LOGS_DIR = PROJECT_ROOT / "logs"
+VOICE_METRICS_PATH = LOGS_DIR / "voice_metrics.jsonl"
+
+
+# ---------- Lazy model singletons ----------
+
+_whisper_model: Any = None
+_whisper_lock = threading.Lock()
+_silero_model: Any = None
+_silero_get_timestamps: Any = None
+_silero_lock = threading.Lock()
+
+
+def _get_whisper_model() -> Any:
+    """Lazy-load faster-whisper ``small`` int8 with the spike-validated
+    ``cpu_threads=4`` (see ``tasks/voice-bench-results.md``)."""
+    global _whisper_model
+    if _whisper_model is not None:
+        return _whisper_model
+    with _whisper_lock:
+        if _whisper_model is not None:
+            return _whisper_model
+        from faster_whisper import WhisperModel
+        _whisper_model = WhisperModel(
+            "small", device="cpu", compute_type="int8", cpu_threads=4,
+            local_files_only=True,
+        )
+        return _whisper_model
+
+
+def _get_silero_vad():
+    """Lazy-load silero-vad. Returns ``(model, get_speech_timestamps)``."""
+    global _silero_model, _silero_get_timestamps
+    if _silero_model is not None:
+        return _silero_model, _silero_get_timestamps
+    with _silero_lock:
+        if _silero_model is not None:
+            return _silero_model, _silero_get_timestamps
+        from silero_vad import get_speech_timestamps, load_silero_vad
+        _silero_model = load_silero_vad()
+        _silero_get_timestamps = get_speech_timestamps
+        return _silero_model, _silero_get_timestamps
+
+
+# ---------- Audio helpers ----------
+
+def _pcm48_stereo_to_16_mono(pcm: bytes) -> np.ndarray:
+    """Discord 48kHz s16le stereo bytes -> 16kHz mono float32 in [-1, 1].
+
+    Cheap downsample: average the two channels, then average every 3
+    samples (48k / 3 = 16k). faster-whisper + silero-vad accept the
+    resulting ``np.float32`` array directly.
+    """
+    if not pcm:
+        return np.zeros(0, dtype=np.float32)
+    samples = np.frombuffer(pcm, dtype=np.int16)
+    if samples.size % 2 != 0:
+        samples = samples[:-1]
+    stereo = samples.reshape(-1, 2)
+    mono = stereo.mean(axis=1).astype(np.float32) / 32768.0
+    if mono.size == 0:
+        return mono
+    trim = (mono.size // 3) * 3
+    if trim == 0:
+        return np.zeros(0, dtype=np.float32)
+    mono = mono[:trim].reshape(-1, 3).mean(axis=1)
+    return mono.astype(np.float32)
+
+
+# ---------- VoiceSession ----------
+
+class VoiceSession:
+    """Per-voice-call state with a single idempotent ``cleanup()``."""
+
+    def __init__(
+        self,
+        *,
+        channel_id: int,
+        guild_id: int,
+        text_channel: Any,
+        voice_client: Any,
+        bot: Any,
+        ttsq: Any,
+        whitelist: Optional[set] = None,
+        record_enabled: bool = False,
+        mirror_enabled: bool = True,
+        transcripts_jsonl_path: Optional[Path] = None,
+        loop: Optional[asyncio.AbstractEventLoop] = None,
+        router_route_message: Optional[Callable] = None,
+    ):
+        self.channel_id = int(channel_id)
+        self.guild_id = int(guild_id)
+        self.text_channel = text_channel
+        self.voice_client = voice_client
+        self.bot = bot
+        self.ttsq = ttsq
+        self.whitelist: set = set(whitelist or set())
+        self.record_enabled = bool(record_enabled)
+        self.mirror_enabled = bool(mirror_enabled)
+        self.transcripts_jsonl_path = transcripts_jsonl_path
+        self.loop = loop
+        # Injection seam so tests can replace router.route_message without
+        # mocking the whole module.
+        if router_route_message is None:
+            from src.router import route_message as _rm
+            self._route_message = _rm
+        else:
+            self._route_message = router_route_message
+
+        self.last_activity_ts = time.monotonic()
+        self._jsonl_fh = None
+        self._lock = threading.Lock()
+        self._cleaned_up = False
+        self._lock_owner_thread: Optional[int] = None
+
+    # ----- context manager -----
+
+    def __enter__(self) -> "VoiceSession":
+        self._lock.acquire()
+        self._lock_owner_thread = threading.get_ident()
+        if self.record_enabled and self.transcripts_jsonl_path is not None:
+            try:
+                self.transcripts_jsonl_path.parent.mkdir(
+                    parents=True, exist_ok=True,
+                )
+                self._jsonl_fh = open(
+                    self.transcripts_jsonl_path, "a",
+                    buffering=1, encoding="utf-8",
+                )
+            except OSError as e:
+                log.warning("voice transcript open failed: %s", e)
+                self._jsonl_fh = None
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb) -> bool:
+        self.cleanup("exit")
+        return False  # never suppress exceptions
+
+    # ----- cleanup (centralized, idempotent) -----
+
+    def cleanup(self, reason: str) -> None:
+        """Single drain path for ALL 5 exit scenarios. Safe to call twice."""
+        if self._cleaned_up:
+            return
+        self._cleaned_up = True
+
+        # 1. Flush or discard JSONL transcript.
+        if self._jsonl_fh is not None:
+            try:
+                self._jsonl_fh.flush()
+                self._jsonl_fh.close()
+            except Exception as e:  # noqa: BLE001
+                log.warning("voice transcript flush failed: %s", e)
+            self._jsonl_fh = None
+        if (not self.record_enabled
+                and self.transcripts_jsonl_path is not None
+                and self.transcripts_jsonl_path.exists()):
+            try:
+                self.transcripts_jsonl_path.unlink()
+            except OSError:
+                pass
+
+        # 2. Restore bot presence (clear Listening activity).
+        if self.bot is not None:
+            try:
+                change = getattr(self.bot, "change_presence", None)
+                if callable(change):
+                    coro = change(activity=None)
+                    if asyncio.iscoroutine(coro):
+                        if self.loop is not None and self.loop.is_running():
+                            asyncio.run_coroutine_threadsafe(coro, self.loop)
+                        else:
+                            # Best-effort: close the coroutine so Python
+                            # doesn't emit "coroutine was never awaited".
+                            coro.close()
+            except Exception as e:  # noqa: BLE001
+                log.warning("voice presence restore failed: %s", e)
+
+        # 3. Tear down the voice client.
+        if self.voice_client is not None:
+            try:
+                self.voice_client.cleanup()
+            except Exception as e:  # noqa: BLE001
+                log.warning("voice_client.cleanup failed: %s", e)
+
+        # 4. Stop the TTS queue worker.
+        if self.ttsq is not None:
+            try:
+                self.ttsq.stop()
+            except Exception as e:  # noqa: BLE001
+                log.warning("ttsq.stop failed: %s", e)
+
+        # 5. Release the session lock (held since __enter__).
+        try:
+            if self._lock.locked():
+                self._lock.release()
+        except RuntimeError:
+            # Released from a different thread than acquired it — already
+            # free for the next caller; nothing to do.
+            pass
+
+        self._log_metric({"event": "cleanup", "reason": reason})
+
+    # ----- segment completion (scheduled from sink) -----
+
+    async def on_segment_done(
+        self,
+        speaker_id: int,
+        text: str,
+        no_speech_prob: float,
+    ) -> None:
+        """Mirror, persist, route to Claude, drive TTS via streaming callback."""
+        if self._cleaned_up:
+            return
+        self.last_activity_ts = time.monotonic()
+        speaker_name = self._resolve_speaker_name(speaker_id)
+
+        # Drop any TTS frames from the previous turn so a new utterance cuts off
+        # stale Echo speech (barge-in) and never mixes with the new response.
+        try:
+            self.ttsq.clear()
+        except Exception as e:  # noqa: BLE001
+            log.warning("ttsq.clear failed: %s", e)
+
+        # 1. Mirror to text channel (one Unicode 🎤 — exception per plan).
+        if self.mirror_enabled and self.text_channel is not None:
+            try:
+                send = getattr(self.text_channel, "send", None)
+                if callable(send):
+                    coro = send(f"\U0001f3a4 {speaker_name}: \"{text}\"")
+                    if asyncio.iscoroutine(coro):
+                        await coro
+            except Exception as e:  # noqa: BLE001
+                log.warning("voice mirror send failed: %s", e)
+
+        # 2. Append to JSONL transcript buffer if recording.
+        if self._jsonl_fh is not None:
+            try:
+                self._jsonl_fh.write(
+                    json.dumps({
+                        "ts": time.time(),
+                        "speaker_id": speaker_id,
+                        "speaker": speaker_name,
+                        "text": text,
+                        "no_speech_prob": no_speech_prob,
+                    }, ensure_ascii=False) + "\n"
+                )
+            except Exception as e:  # noqa: BLE001
+                log.warning("voice transcript write failed: %s", e)
+
+        block_count = [0]
+
+        def voice_stream_callback(block: str) -> None:
+            """Called once per Claude streamed text block — pushes to TTS."""
+            block_count[0] += 1
+            log.info("voice stream block #%d (%d chars): %r",
+                     block_count[0], len(block or ""), (block or "")[:80])
+            try:
+                self.ttsq.push_text(block)
+            except Exception as e:  # noqa: BLE001
+                log.warning("ttsq.push_text failed: %s", e)
+
+        # Dispatch to Claude. send_message is sync subprocess, run on
+        # a worker thread so the loop stays responsive for mirror/TTS.
+        try:
+            await asyncio.to_thread(
+                self._route_message,
+                str(self.channel_id),
+                str(speaker_id),
+                text,
+                None,                       # model
+                voice_stream_callback,      # on_text
+                "discord-voice",            # adapter_name
+            )
+        except Exception as e:  # noqa: BLE001
+            log.error("route_message voice path failed: %s", e)
+
+    # ----- helpers -----
+
+    def _resolve_speaker_name(self, speaker_id: int) -> str:
+        """Best-effort display name lookup via the bot user cache."""
+        try:
+            if self.bot is not None and hasattr(self.bot, "get_user"):
+                user = self.bot.get_user(speaker_id)
+                if user is not None:
+                    name = getattr(user, "display_name", None) or getattr(
+                        user, "name", None,
+                    )
+                    if name:
+                        return str(name)
+        except Exception:  # noqa: BLE001
+            pass
+        return str(speaker_id)
+
+    def _log_metric(self, payload: dict) -> None:
+        """Append a structured event to ``logs/voice_metrics.jsonl``."""
+        event = {"ts": time.time(), "channel_id": self.channel_id, **payload}
+        try:
+            LOGS_DIR.mkdir(parents=True, exist_ok=True)
+            with open(VOICE_METRICS_PATH, "a", buffering=1, encoding="utf-8") as f:
+                f.write(json.dumps(event, ensure_ascii=False) + "\n")
+        except OSError:
+            pass
+
+
+# ---------- EchoVoiceSink ----------
+
+class EchoVoiceSink(AudioSink):
+    """PCM-in sink: per-user 20ms buffer -> 100ms VAD windows -> 800ms
+    silence triggers Whisper STT -> schedules ``on_segment_done`` on the
+    session loop.
+
+    Lives in the voice_recv reader thread; uses ``threading`` primitives
+    only (no asyncio in the hot path).
+    """
+
+    def __init__(self, session: VoiceSession, bot_user_id: int):
+        super().__init__()
+        self.session = session
+        self.bot_user_id = int(bot_user_id) if bot_user_id is not None else 0
+        self.whitelist: set = set(session.whitelist or set())
+        self._user_buffers: dict[int, bytearray] = {}
+        self._packet_accum: dict[int, bytearray] = {}
+        self._last_speech_ts: dict[int, float] = {}
+        self._has_speech: dict[int, bool] = {}
+        self._sink_lock = threading.Lock()
+        # Diagnostics: log once-per-user when packets first arrive and when
+        # VAD first detects speech. Cheap, but tells us exactly where the
+        # chain breaks when "I spoke but Echo heard nothing" happens.
+        self._first_packet_logged: set[int] = set()
+        self._first_speech_logged: set[int] = set()
+        # Background poller that triggers the silence flush even when Discord
+        # DTX stops delivering RTP packets after the user stops speaking. Without
+        # this, sink.write would stop firing and STT would never run on the
+        # final utterance.
+        self._poller_stop = threading.Event()
+        self._poller_thread = threading.Thread(
+            target=self._silence_flush_poller,
+            name="echo-voice-flush-poller",
+            daemon=True,
+        )
+        self._poller_thread.start()
+
+    def wants_opus(self) -> bool:
+        return False
+
+    def cleanup(self) -> None:
+        self._poller_stop.set()
+        with self._sink_lock:
+            self._user_buffers.clear()
+            self._packet_accum.clear()
+            self._last_speech_ts.clear()
+            self._has_speech.clear()
+
+    def write(self, user, voice_data: VoiceData) -> None:
+        # ---- FIRST GUARD (LOAD-BEARING): bot's own voice ---------------
+        if user is None:
+            return
+        uid = int(getattr(user, "id", 0) or 0)
+        if uid == 0:
+            return
+        if uid == self.bot_user_id:
+            return
+
+        # ---- SECOND GUARD: whitelist filter ----------------------------
+        if self.whitelist and uid not in self.whitelist:
+            return
+
+        pcm = getattr(voice_data, "pcm", None)
+        if not pcm:
+            return
+
+        if uid not in self._first_packet_logged:
+            self._first_packet_logged.add(uid)
+            log.info("voice sink: first PCM packet from user %s (%d bytes)", uid, len(pcm))
+
+        window_pcm: Optional[bytes] = None
+        pcm_for_stt: Optional[bytes] = None
+
+        try:
+            with self._sink_lock:
+                buf = self._user_buffers.setdefault(uid, bytearray())
+                accum = self._packet_accum.setdefault(uid, bytearray())
+                buf.extend(pcm)
+                accum.extend(pcm)
+                if len(accum) >= VAD_WINDOW_BYTES:
+                    window_pcm = bytes(accum[:VAD_WINDOW_BYTES])
+                    del accum[:VAD_WINDOW_BYTES]
+
+            if window_pcm is not None:
+                if self._vad_detects_speech(window_pcm):
+                    if uid not in self._first_speech_logged:
+                        self._first_speech_logged.add(uid)
+                        log.info("voice sink: VAD detected speech from user %s", uid)
+                    with self._sink_lock:
+                        self._last_speech_ts[uid] = time.monotonic()
+                        self._has_speech[uid] = True
+
+            pcm_for_stt = self._take_flushable_pcm(uid)
+            if pcm_for_stt:
+                self._flush_to_stt(uid, pcm_for_stt)
+        except Exception as e:  # noqa: BLE001
+            log.warning("EchoVoiceSink.write failed: %s", e)
+
+    def _take_flushable_pcm(self, uid: int) -> Optional[bytes]:
+        """If user `uid` has buffered speech that's been silent ≥SILENCE_FLUSH_MS,
+        consume the buffer and return it. Otherwise return None."""
+        with self._sink_lock:
+            if not self._has_speech.get(uid):
+                return None
+            last = self._last_speech_ts.get(uid, 0.0)
+            silence_ms = (time.monotonic() - last) * 1000.0
+            if silence_ms < SILENCE_FLUSH_MS:
+                return None
+            pcm = bytes(self._user_buffers.get(uid, b""))
+            self._user_buffers[uid] = bytearray()
+            self._packet_accum[uid] = bytearray()
+            self._has_speech[uid] = False
+            return pcm if pcm else None
+
+    def _silence_flush_poller(self) -> None:
+        """Background tick: Discord DTX stops sending RTP packets when the user
+        goes silent, so the inline flush check in `write()` never fires for the
+        last utterance. Poll every 200ms so the trailing audio actually reaches
+        Whisper."""
+        while not self._poller_stop.wait(0.2):
+            try:
+                with self._sink_lock:
+                    pending = [uid for uid, has in self._has_speech.items() if has]
+                for uid in pending:
+                    pcm = self._take_flushable_pcm(uid)
+                    if pcm:
+                        self._flush_to_stt(uid, pcm)
+            except Exception as e:  # noqa: BLE001
+                log.warning("silence flush poller iter failed: %s", e)
+
+    # ----- VAD -----
+
+    def _vad_detects_speech(self, pcm48_stereo: bytes) -> bool:
+        """Run silero-vad on a 100ms window. silero-vad v5+ requires exactly
+        512 samples per call at 16kHz, so we slice the window into 512-sample
+        chunks and return True if any chunk crosses the threshold."""
+        try:
+            mono16 = _pcm48_stereo_to_16_mono(pcm48_stereo)
+            if mono16.size == 0:
+                return False
+            try:
+                import torch
+            except ImportError:
+                rms = float(np.sqrt(np.mean(mono16.astype(np.float64) ** 2)))
+                return rms > 0.02
+            model, _ = _get_silero_vad()
+            chunk = 512  # silero-vad v5+ hard requirement at 16kHz
+            max_prob = 0.0
+            with torch.no_grad():
+                for start in range(0, mono16.size - chunk + 1, chunk):
+                    seg = mono16[start:start + chunk]
+                    p = float(model(torch.from_numpy(seg), SAMPLE_RATE_WHISPER).item())
+                    if p > max_prob:
+                        max_prob = p
+                    if p >= VAD_THRESHOLD:
+                        return True
+            return False
+        except Exception as e:  # noqa: BLE001
+            log.debug("VAD inference failed: %s", e)
+            return False
+
+    # ----- STT flush -----
+
+    def _flush_to_stt(self, user_id: int, pcm48_stereo: bytes) -> None:
+        """Downsample, Whisper-transcribe RO, drop hallucinations, dispatch."""
+        try:
+            mono16 = _pcm48_stereo_to_16_mono(pcm48_stereo)
+            if mono16.size == 0:
+                return
+            model = _get_whisper_model()
+            segments, _info = model.transcribe(
+                mono16, language="ro", beam_size=5,
+                initial_prompt=(
+                    "Echo Core, asistent personal AI românesc al lui Marius. "
+                    "Conversație colocvială în română."
+                ),
+                condition_on_previous_text=False,
+            )
+            text_parts: list[str] = []
+            worst_no_speech = 0.0
+            for seg in segments:
+                no_sp = float(getattr(seg, "no_speech_prob", 0.0) or 0.0)
+                if no_sp > worst_no_speech:
+                    worst_no_speech = no_sp
+                if no_sp > NO_SPEECH_DROP_THRESHOLD:
+                    continue
+                seg_text = (getattr(seg, "text", "") or "").strip()
+                if seg_text:
+                    text_parts.append(seg_text)
+            if not text_parts:
+                return
+            text = " ".join(text_parts).strip()
+            if not text:
+                return
+            self._schedule_segment_done(user_id, text, worst_no_speech)
+        except Exception as e:  # noqa: BLE001
+            log.warning("Whisper transcribe failed: %s", e)
+
+    def _schedule_segment_done(
+        self, user_id: int, text: str, no_speech_prob: float,
+    ) -> None:
+        loop = self.session.loop
+        if loop is None or not loop.is_running():
+            log.debug("voice session loop missing — dropping segment")
+            return
+        try:
+            asyncio.run_coroutine_threadsafe(
+                self.session.on_segment_done(user_id, text, no_speech_prob),
+                loop,
+            )
+        except Exception as e:  # noqa: BLE001
+            log.warning("voice segment dispatch failed: %s", e)
+
+
+__all__ = [
+    "VoiceSession",
+    "EchoVoiceSink",
+    "SILENCE_FLUSH_MS",
+    "VAD_THRESHOLD",
+    "VAD_WINDOW_MS",
+    "NO_SPEECH_DROP_THRESHOLD",
+]

src/voice/tts_stream.py

@@ -0,0 +1,333 @@
+"""Streaming TTS with clause-level chunking for Discord voice mode.
+
+A worker thread consumes text -> produces 20ms PCM frames on a queue.Queue.
+``EchoStreamingAudioSource`` pulls frames into Discord's audio thread so a
+single ``voice_client.play()`` call lasts the whole turn (eliminates the
+RTP gap between successive ``play()`` calls and the race with barge-in
+``stop()``). See plan: src/voice/tts_stream.py (Pas 6 / Lane TTS),
+Engineering decisions #6, #8, #15.
+"""
+from __future__ import annotations
+
+import io
+import logging
+import queue
+import re
+import subprocess
+import threading
+import wave
+from pathlib import Path
+from typing import Iterator, List, Optional
+
+import discord
+
+from src.voice.normalize import normalize_for_tts
+from tools.tts import synthesize
+
+
+log = logging.getLogger(__name__)
+
+
+# Discord wants 20ms of 16-bit 48kHz stereo PCM per frame.
+# 48000 Hz * 0.020 s * 2 channels * 2 bytes = 3840 bytes.
+FRAME_BYTES = 3840
+TARGET_SAMPLE_RATE = 48000
+TARGET_CHANNELS = 2
+TARGET_SAMPLE_WIDTH = 2
+
+# Sentinel pushed onto the text queue to ask the worker to exit cleanly.
+_POISON = object()
+
+
+# ---------- Clause segmentation ----------
+
+# Split at Romanian sentence punctuation followed by whitespace. The
+# trailing whitespace requirement protects mid-number (1.000), mid-decimal
+# (12.5), and mid-abbreviation (M.D.) tokens, since none of those have a
+# space right after the inner punctuation.
+_CLAUSE_SPLIT = re.compile(r'(?<=[,;:.!?])\s+')
+
+
+def clause_segments(text: str, min_words: int = 8) -> Iterator[str]:
+    """Yield text in clause-sized chunks for streaming TTS.
+
+    Splits at ``, ; : . ! ?`` boundaries (only when the punctuation is
+    followed by whitespace, so numbers / decimals / abbreviations stay
+    intact). Short clauses are buffered and merged with the next one
+    until the accumulated chunk has at least ``min_words`` words. The
+    final remainder is always yielded, even if it's shorter than
+    ``min_words`` -- otherwise the tail of the response would never
+    reach the TTS.
+    """
+    if text is None:
+        return
+    text = text.strip()
+    if not text:
+        return
+    pieces = [p.strip() for p in _CLAUSE_SPLIT.split(text) if p and p.strip()]
+    if not pieces:
+        return
+    buffer = ''
+    for clause in pieces:
+        buffer = (buffer + ' ' + clause).strip() if buffer else clause
+        if len(buffer.split()) >= min_words:
+            yield buffer
+            buffer = ''
+    if buffer:
+        yield buffer
+
+
+# ---------- WAV -> PCM frame conversion ----------
+
+def _ffmpeg_resample(wav_bytes: bytes) -> bytes:
+    """Convert any WAV payload to raw 48kHz stereo s16le PCM via ffmpeg.
+
+    ffmpeg is already an Echo Core hard dependency (heartbeat, video
+    transcription). Using a stdin/stdout pipe keeps the synth tempfile
+    short-lived and avoids extra disk traffic.
+    """
+    proc = subprocess.run(
+        [
+            'ffmpeg', '-hide_banner', '-loglevel', 'error',
+            '-i', 'pipe:0',
+            '-f', 's16le',
+            '-ar', str(TARGET_SAMPLE_RATE),
+            '-ac', str(TARGET_CHANNELS),
+            '-acodec', 'pcm_s16le',
+            'pipe:1',
+        ],
+        input=wav_bytes,
+        capture_output=True,
+        check=False,
+    )
+    if proc.returncode != 0:
+        err = proc.stderr.decode('utf-8', errors='replace')[:200]
+        raise RuntimeError(f"ffmpeg resample failed (rc={proc.returncode}): {err}")
+    return proc.stdout
+
+
+def _is_target_format(wav_bytes: bytes) -> bool:
+    """Quick check whether the WAV already matches Discord's PCM format."""
+    try:
+        with wave.open(io.BytesIO(wav_bytes), 'rb') as w:
+            return (
+                w.getframerate() == TARGET_SAMPLE_RATE
+                and w.getnchannels() == TARGET_CHANNELS
+                and w.getsampwidth() == TARGET_SAMPLE_WIDTH
+                and w.getcomptype() == 'NONE'
+            )
+    except (wave.Error, EOFError):
+        return False
+
+
+def _extract_pcm_native(wav_bytes: bytes) -> bytes:
+    """Strip the WAV header and return raw PCM (target format assumed)."""
+    with wave.open(io.BytesIO(wav_bytes), 'rb') as w:
+        return w.readframes(w.getnframes())
+
+
+def wav_to_pcm_20ms_frames(wav_bytes: bytes) -> List[bytes]:
+    """Parse a WAV blob, normalize to 48kHz s16le stereo, slice into 20ms frames.
+
+    The final frame is zero-padded to a full 3840 bytes so Discord's audio
+    thread always reads whole frames. Empty input yields an empty list.
+    """
+    if not wav_bytes:
+        return []
+    pcm = _extract_pcm_native(wav_bytes) if _is_target_format(wav_bytes) else _ffmpeg_resample(wav_bytes)
+    if not pcm:
+        return []
+    frames: List[bytes] = []
+    for offset in range(0, len(pcm), FRAME_BYTES):
+        chunk = pcm[offset:offset + FRAME_BYTES]
+        if len(chunk) < FRAME_BYTES:
+            chunk = chunk + b'\x00' * (FRAME_BYTES - len(chunk))
+        frames.append(chunk)
+    return frames
+
+
+# ---------- TTS worker queue ----------
+
+class TTSQueue:
+    """Worker thread: text in -> 20ms PCM frames out.
+
+    Usage::
+
+        ttsq = TTSQueue(voice_id="M2", lang="ro")
+        ttsq.start()
+        ttsq.push_text("salut Marius, ce mai faci?")
+        voice_client.play(EchoStreamingAudioSource(ttsq))
+        # ... barge-in detected:
+        ttsq.clear()
+        # ... session over:
+        ttsq.stop()
+    """
+
+    def __init__(self, voice_id: str = "M2", lang: str = "ro"):
+        self.voice_id = voice_id
+        self.lang = lang
+        self._text_queue: queue.Queue = queue.Queue()
+        self._pcm_queue: queue.Queue = queue.Queue()
+        self._worker_thread: Optional[threading.Thread] = None
+        self._stop_event = threading.Event()
+
+    # --- lifecycle ---
+
+    def start(self) -> None:
+        if self._worker_thread is not None and self._worker_thread.is_alive():
+            return
+        self._stop_event.clear()
+        self._worker_thread = threading.Thread(
+            target=self._worker_loop,
+            name=f"tts-worker-{self.voice_id}",
+            daemon=True,
+        )
+        self._worker_thread.start()
+
+    def stop(self) -> None:
+        """Signal the worker to exit, drain queues, join (timeout 5s)."""
+        self._stop_event.set()
+        # Wake the worker if it's blocked on get(timeout=...).
+        self._text_queue.put(_POISON)
+        thread = self._worker_thread
+        if thread is not None:
+            thread.join(timeout=5.0)
+        self._worker_thread = None
+        self._drain(self._text_queue)
+        self._drain(self._pcm_queue)
+
+    # --- producer side ---
+
+    def push_text(self, text: str) -> None:
+        """Normalize, segment into clauses, enqueue each clause for synthesis."""
+        if not text:
+            return
+        cleaned = normalize_for_tts(text)
+        n = 0
+        for clause in clause_segments(cleaned):
+            clause = clause.strip()
+            if clause:
+                self._text_queue.put(clause)
+                n += 1
+        log.info("ttsq.push_text: input %d chars → %d clauses queued", len(text), n)
+
+    def clear(self) -> None:
+        """Drop everything pending (used for barge-in)."""
+        self._drain(self._text_queue)
+        self._drain(self._pcm_queue)
+
+    def is_empty(self) -> bool:
+        return self._text_queue.empty() and self._pcm_queue.empty()
+
+    # --- consumer side (called by EchoStreamingAudioSource) ---
+
+    def get_frame_nowait(self) -> Optional[bytes]:
+        """Return the next PCM frame if available, else None — no blocking.
+
+        Blocking inside the player's read() loop wrecks Discord's 20ms cadence
+        and the client interprets the stream as stuttering / out-of-order.
+        """
+        try:
+            return self._pcm_queue.get_nowait()
+        except queue.Empty:
+            return None
+
+    # --- internals ---
+
+    @staticmethod
+    def _drain(q: queue.Queue) -> None:
+        while True:
+            try:
+                q.get_nowait()
+            except queue.Empty:
+                return
+
+    def _worker_loop(self) -> None:
+        while not self._stop_event.is_set():
+            try:
+                item = self._text_queue.get(timeout=0.1)
+            except queue.Empty:
+                continue
+            if item is _POISON:
+                break
+            if not isinstance(item, str):
+                continue
+            preview = item[:60]
+            try:
+                result = synthesize(item, voice=self.voice_id, lang=self.lang)
+            except Exception as e:
+                log.warning("TTS synth raised for %r: %s", preview, e)
+                continue
+            if not result.get('ok'):
+                log.warning("TTS synth not ok for %r: %s", preview, result.get('error'))
+                continue
+            path = result.get('path')
+            if not path:
+                log.warning("TTS synth ok but no path for %r", preview)
+                continue
+            wav_bytes = b''
+            try:
+                wav_bytes = Path(path).read_bytes()
+            except OSError as e:
+                log.warning("TTS WAV read failed for %r: %s", preview, e)
+            finally:
+                try:
+                    Path(path).unlink(missing_ok=True)
+                except OSError:
+                    pass
+            if not wav_bytes:
+                continue
+            try:
+                frames = wav_to_pcm_20ms_frames(wav_bytes)
+            except RuntimeError as e:
+                log.warning("TTS WAV-to-PCM failed for %r: %s", preview, e)
+                continue
+            if not frames:
+                log.warning("TTS WAV-to-PCM produced 0 frames for %r", preview)
+                continue
+            for frame in frames:
+                if self._stop_event.is_set():
+                    return
+                self._pcm_queue.put(frame)
+            log.info("TTS pushed %d frames (%.1fs) for %r",
+                     len(frames), len(frames) * 0.02, preview)
+
+
+# ---------- Discord audio source ----------
+
+class EchoStreamingAudioSource(discord.AudioSource):
+    """Pull PCM frames from a ``TTSQueue`` into Discord's audio thread.
+
+    A single ``voice_client.play(EchoStreamingAudioSource(ttsq))`` call
+    spans the whole session. When the TTS queue is empty, ``read()``
+    returns a 20ms silence frame to keep the player alive — otherwise
+    Discord would interpret an empty return as end-of-stream and stop
+    the player, so real TTS frames pushed later would be silently
+    discarded. The player is explicitly terminated only via
+    ``cleanup()`` (called on voice session teardown).
+    """
+
+    # 20ms of s16le stereo at 48kHz silence (960 samples × 2 channels × 2 bytes).
+    _SILENCE_FRAME = b'\x00' * (960 * 2 * 2)
+
+    def __init__(self, ttsq: TTSQueue):
+        self._ttsq = ttsq
+        self._closed = False
+
+    def read(self) -> bytes:
+        if self._closed:
+            return b''
+        frame = self._ttsq.get_frame_nowait()
+        if frame is None:
+            return self._SILENCE_FRAME
+        return frame
+
+    def is_opus(self) -> bool:
+        return False
+
+    def cleanup(self) -> None:
+        self._closed = True
+        try:
+            self._ttsq.clear()
+        except Exception:
+            pass

tasks/lessons.md

@@ -17,6 +17,13 @@ Lecții capturate din corectările lui Marius. Citește acest fișier la începu
 
 <!-- Lecțiile se adaugă mai jos, cele mai noi sus. -->
 
+## Supertonic rejectează ghilimelele curly (Unicode) cu HTTP 500
+**Data:** 2026-05-27
+**Context:** Marius a dat o comandă audio pe Discord cu un URL, iar răspunsul lui Claude conținea `„foo"` (ghilimele românești curly). Supertonic a returnat `HTTP 500: synthesis failed: Found 1 unsupported character(s): ['„']` și răspunsul nu s-a mai auzit. Fără retry logic vizibil în UX — pur și simplu tace.
+**Greșeala:** Am presupus că `normalize_for_tts` produce text deja "TTS-safe" pentru Supertonic. În realitate `strip_markdown` păstrează ghilimelele Unicode (`„` U+201E, `"` U+201D, `—` U+2014, `…` U+2026, etc.) pe care Supertonic le refuză.
+**Regula:** Înainte de orice apel HTTP la Supertonic, **sanitizează punctuația Unicode** la echivalentele ASCII (`„` `"` `"` → `"`, `'` `'` `‚` → `'`, `–` `—` → `-`, `…` → `...`, `«` `»` → `"`). Funcția `sanitize_punctuation` în `src/voice/normalize.py` face asta și e apelată chiar după `strip_markdown` în pipeline. Dacă apar caractere noi care crapă Supertonic (ex: simboluri matematice, săgeți), adaugă-le în `_TTS_PUNCT_MAP`.
+**Când se aplică:** Orice cod care trimite text la Supertonic (`tools/tts.py`, `src/voice/tts_stream.py`). Inclusiv testare manuală cu `curl` — folosește text românesc realistic (include `„foo"`, em-dash `—`, ellipsis `…`).
+
 ## Mai multe threads ≠ mai rapid — fitează `cpu_threads` pe physical cores, nu logical
 **Data:** 2026-05-27
 **Context:** Benchmark `tools/voice_bench.py` pentru faster-whisper `small` int8 pe i7-6700T (4 physical / 8 logical cores). Marius a urcat VM-ul de la 2 → 4 → 6 cores online, așteptând că mai multe = mai rapid.

tests/test_claude_session_mutex.py

@@ -0,0 +1,307 @@
+"""Regression-critical tests for per-channel mutex in src/claude_session.py.
+
+Three scenarios from the eng-review test plan (2026-05-27):
+
+  1. Concurrent `send_message` calls on the SAME channel_id serialize —
+     the second waits for the first to finish before its subprocess runs.
+  2. Concurrent `send_message` calls on DIFFERENT channel_ids run in parallel
+     — independent channels never block each other.
+  3. Acquisition contract is documented and consistent: the lock is acquired
+     blocking (no acquire timeout), which means a hung subprocess on
+     channel X delays subsequent X messages but never X' (X != X'). This
+     test pins that behavior so future refactors must preserve it.
+
+The mutex is `threading.Lock`, NOT `asyncio.Lock`, because `send_message`
+is a sync function typically dispatched via `asyncio.to_thread` from
+async adapters. asyncio.Lock would serialize coroutines only — not the
+subprocess invocation. See plan section "Engineering decisions" #2.
+"""
+
+import json
+import threading
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from unittest.mock import patch
+
+import pytest
+
+from src import claude_session
+from src.claude_session import (
+    _get_session_lock,
+    _session_locks,
+    send_message,
+)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture(autouse=True)
+def _clear_session_locks():
+    """Each test starts with a fresh lock map so we don't share state."""
+    _session_locks.clear()
+    yield
+    _session_locks.clear()
+
+
+@pytest.fixture
+def temp_sessions(tmp_path, monkeypatch):
+    """Isolated active.json per test — keeps real session state untouched."""
+    sessions_dir = tmp_path / "sessions"
+    sessions_dir.mkdir()
+    sf = sessions_dir / "active.json"
+    sf.write_text("{}")
+    monkeypatch.setattr(claude_session, "SESSIONS_DIR", sessions_dir)
+    monkeypatch.setattr(claude_session, "_SESSIONS_FILE", sf)
+    return sf
+
+
+def _slow_run_claude(sleep_seconds: float, in_critical: threading.Event,
+                     concurrent_seen: threading.Event):
+    """Build a fake `_run_claude` that signals when inside the critical section.
+
+    The fake holds the simulated subprocess for `sleep_seconds`. Any other
+    invocation that overlaps will set `concurrent_seen` — the mutex test
+    asserts this NEVER happens for the same channel_id.
+    """
+    state = {"active": 0, "lock": threading.Lock()}
+
+    def fake(cmd, timeout, on_text=None, cwd=None):
+        with state["lock"]:
+            state["active"] += 1
+            if state["active"] > 1:
+                concurrent_seen.set()
+        in_critical.set()
+        time.sleep(sleep_seconds)
+        with state["lock"]:
+            state["active"] -= 1
+        return {
+            "result": "Hello from Claude!",
+            "session_id": "sess-abc-123",
+            "usage": {"input_tokens": 10, "output_tokens": 5},
+            "total_cost_usd": 0.001,
+            "cost_usd": 0.001,
+            "duration_ms": int(sleep_seconds * 1000),
+            "num_turns": 1,
+            "intermediate_count": 0,
+            "subtype": "success",
+            "is_error": False,
+        }
+
+    return fake
+
+
+# ---------------------------------------------------------------------------
+# Scenario 1 — same channel serializes
+# ---------------------------------------------------------------------------
+
+
+class TestSameChannelSerializes:
+    def test_two_concurrent_calls_same_channel_run_one_at_a_time(
+        self, temp_sessions
+    ):
+        """Two parallel send_message on the SAME channel_id never overlap.
+
+        We instrument `_run_claude` to signal whenever more than one
+        invocation is concurrently inside it. The mutex MUST prevent that.
+        """
+        in_critical = threading.Event()
+        concurrent_seen = threading.Event()
+        slow = _slow_run_claude(0.25, in_critical, concurrent_seen)
+
+        with patch.object(claude_session, "_run_claude", side_effect=slow):
+            start = time.monotonic()
+            with ThreadPoolExecutor(max_workers=2) as pool:
+                futures = [
+                    pool.submit(send_message, "ch-same", f"msg-{i}")
+                    for i in range(2)
+                ]
+                results = [f.result(timeout=10) for f in futures]
+            elapsed = time.monotonic() - start
+
+        assert not concurrent_seen.is_set(), (
+            "Two send_message calls on the same channel ran concurrently — "
+            "mutex did not serialize them."
+        )
+        assert all(r == "Hello from Claude!" for r in results)
+        # Two serial 0.25s subprocesses must take at least ~0.5s total
+        # (we allow a generous floor — schedulers can be slow).
+        assert elapsed >= 0.45, f"Expected serialized ~0.5s, got {elapsed:.3f}s"
+
+    def test_lock_is_reentrant_per_channel_dict(self, temp_sessions):
+        """`_get_session_lock` returns the SAME lock object for the same channel."""
+        lock_a1 = _get_session_lock("channel-A")
+        lock_a2 = _get_session_lock("channel-A")
+        lock_b = _get_session_lock("channel-B")
+        assert lock_a1 is lock_a2
+        assert lock_a1 is not lock_b
+
+
+# ---------------------------------------------------------------------------
+# Scenario 2 — different channels parallel
+# ---------------------------------------------------------------------------
+
+
+class TestDifferentChannelsParallel:
+    def test_two_concurrent_calls_different_channels_run_in_parallel(
+        self, temp_sessions
+    ):
+        """Different channels MUST NOT block each other.
+
+        We measure elapsed wall-clock: two 0.4s subprocesses on different
+        channels should finish in ~0.4s (parallel), NOT ~0.8s (serialized).
+        """
+        in_critical = threading.Event()
+        # `concurrent_seen` is OK to fire here — we WANT them to overlap.
+        concurrent_seen = threading.Event()
+        slow = _slow_run_claude(0.4, in_critical, concurrent_seen)
+
+        with patch.object(claude_session, "_run_claude", side_effect=slow):
+            start = time.monotonic()
+            with ThreadPoolExecutor(max_workers=2) as pool:
+                f1 = pool.submit(send_message, "ch-A", "msg-A")
+                f2 = pool.submit(send_message, "ch-B", "msg-B")
+                results = [f1.result(timeout=10), f2.result(timeout=10)]
+            elapsed = time.monotonic() - start
+
+        assert all(r == "Hello from Claude!" for r in results)
+        # Parallel execution: total time should be close to 0.4s, well under
+        # 0.7s (would mean serialization). 0.65s ceiling allows for GIL +
+        # scheduler jitter on a busy test box.
+        assert elapsed < 0.65, (
+            f"Different channels appear serialized: elapsed {elapsed:.3f}s "
+            f"(expected ~0.4s parallel, <0.65s ceiling)"
+        )
+        assert concurrent_seen.is_set(), (
+            "Different channels did not overlap — mutex is too coarse "
+            "(should be per-channel, not global)."
+        )
+
+    def test_three_channels_all_overlap(self, temp_sessions):
+        """Stress: three concurrent channels all run in parallel."""
+        in_critical = threading.Event()
+        concurrent_seen = threading.Event()
+        slow = _slow_run_claude(0.3, in_critical, concurrent_seen)
+
+        with patch.object(claude_session, "_run_claude", side_effect=slow):
+            start = time.monotonic()
+            with ThreadPoolExecutor(max_workers=3) as pool:
+                futures = [
+                    pool.submit(send_message, f"ch-{i}", f"msg-{i}")
+                    for i in range(3)
+                ]
+                for f in as_completed(futures, timeout=10):
+                    assert f.result() == "Hello from Claude!"
+            elapsed = time.monotonic() - start
+
+        # 3 × 0.3s in parallel ≈ 0.3s; serial would be ~0.9s.
+        assert elapsed < 0.6, (
+            f"Three channels serialized: {elapsed:.3f}s (expected <0.6s)"
+        )
+
+
+# ---------------------------------------------------------------------------
+# Scenario 3 — acquisition behavior documented and consistent
+# ---------------------------------------------------------------------------
+
+
+class TestAcquisitionBehavior:
+    """Pin the chosen acquisition policy: blocking, no timeout.
+
+    Project style is to bound subprocess execution via `timeout` (default
+    5 min) rather than fail-fast on lock acquire. Reasons:
+
+    - Adapter callers (Discord/Telegram/voice) already serialize work via
+      asyncio.to_thread; queue depth is naturally bounded.
+    - A non-blocking acquire would surface a timing error to the user
+      ("busy, try again") for an entirely transient and self-resolving
+      condition. Blocking gives FIFO-ish ordering with simple semantics.
+    - If a subprocess truly hangs past `timeout`, _run_claude raises
+      TimeoutError → the held lock releases (via `with`) → queued
+      callers proceed.
+
+    This test pins that: a second caller waits and eventually proceeds; it
+    does not raise an exception on contention.
+    """
+
+    def test_contested_acquire_blocks_then_proceeds(self, temp_sessions):
+        in_critical = threading.Event()
+        concurrent_seen = threading.Event()
+        slow = _slow_run_claude(0.3, in_critical, concurrent_seen)
+
+        results: list[str | BaseException] = []
+
+        def run(label: str):
+            try:
+                results.append(send_message("ch-contend", label))
+            except BaseException as e:
+                results.append(e)
+
+        with patch.object(claude_session, "_run_claude", side_effect=slow):
+            t1 = threading.Thread(target=run, args=("first",))
+            t1.start()
+            # Wait until the first call is inside the critical section so
+            # the second is GUARANTEED to contend on the lock.
+            assert in_critical.wait(timeout=2.0), "first call never entered"
+            in_critical.clear()
+            t2 = threading.Thread(target=run, args=("second",))
+            t2.start()
+            t1.join(timeout=5.0)
+            t2.join(timeout=5.0)
+
+        assert len(results) == 2
+        # Both must return the canned response — no exception, no error.
+        assert all(r == "Hello from Claude!" for r in results), (
+            f"Contended acquire surfaced an error instead of blocking: {results}"
+        )
+        # Critical-section overlap check: contended calls MUST serialize.
+        assert not concurrent_seen.is_set(), (
+            "Contended same-channel calls ran concurrently — mutex broken."
+        )
+
+    def test_lock_released_on_subprocess_exception(self, temp_sessions):
+        """If `_run_claude` raises, the lock MUST be released so the next
+        caller can proceed (otherwise a single error deadlocks the channel
+        forever)."""
+
+        call_count = {"n": 0}
+
+        def flaky(cmd, timeout, on_text=None, cwd=None):
+            call_count["n"] += 1
+            if call_count["n"] == 1:
+                raise RuntimeError("simulated subprocess crash")
+            return {
+                "result": "Hello from Claude!",
+                "session_id": "sess-abc-123",
+                "usage": {"input_tokens": 10, "output_tokens": 5},
+                "total_cost_usd": 0.001,
+                "cost_usd": 0.001,
+                "duration_ms": 50,
+                "num_turns": 1,
+                "intermediate_count": 0,
+                "subtype": "success",
+                "is_error": False,
+            }
+
+        with patch.object(claude_session, "_run_claude", side_effect=flaky):
+            with pytest.raises(RuntimeError, match="simulated subprocess crash"):
+                send_message("ch-recover", "first")
+
+            # Second call MUST acquire the lock (proves the first released it).
+            # We use a short timeout via a thread so a deadlock would fail loudly.
+            done = threading.Event()
+            result_box: list[str] = []
+
+            def second():
+                result_box.append(send_message("ch-recover", "second"))
+                done.set()
+
+            t = threading.Thread(target=second)
+            t.start()
+            assert done.wait(timeout=3.0), (
+                "Second call deadlocked — lock was not released on exception."
+            )
+            t.join(timeout=1.0)
+            assert result_box == ["Hello from Claude!"]

tests/test_voice_adapter_contract.py

@@ -0,0 +1,222 @@
+# -*- coding: utf-8 -*-
+"""Contract test for `src/voice/_discord_voice_adapter.py`.
+
+Purpose: catch drift when the vendored `discord-ext-voice-recv` is upgraded.
+If upstream renames/removes a method we depend on, this test fails LOUDLY
+before any downstream code breaks at runtime in a Discord voice call.
+
+Per VENDOR_INFO.md: this test MUST PASS after every vendor upgrade.
+
+Plain `import` + `hasattr` / `callable` checks — no mocks. We're verifying
+the SHAPE of the API surface, not behavior.
+"""
+
+from __future__ import annotations
+
+import inspect
+
+import pytest
+
+
+# --- Adapter re-exports import cleanly --------------------------------------
+
+
+def test_adapter_exports_voice_receive_client():
+    from src.voice._discord_voice_adapter import VoiceReceiveClient
+
+    assert VoiceReceiveClient is not None
+    assert inspect.isclass(VoiceReceiveClient)
+
+
+def test_adapter_exports_audio_sink():
+    from src.voice._discord_voice_adapter import AudioSink
+
+    assert AudioSink is not None
+    assert inspect.isclass(AudioSink)
+
+
+def test_adapter_exports_voice_data():
+    from src.voice._discord_voice_adapter import VoiceData
+
+    assert VoiceData is not None
+    assert inspect.isclass(VoiceData)
+
+
+def test_adapter_exports_connect_helper():
+    from src.voice._discord_voice_adapter import connect_voice
+
+    assert callable(connect_voice)
+    assert inspect.iscoroutinefunction(connect_voice)
+
+
+# --- Re-exports point at the real vendored classes (no accidental shadowing) -
+
+
+def test_voice_receive_client_is_voice_recv_client():
+    from discord.ext import voice_recv
+
+    from src.voice._discord_voice_adapter import VoiceReceiveClient
+
+    assert VoiceReceiveClient is voice_recv.VoiceRecvClient
+
+
+def test_audio_sink_is_voice_recv_audio_sink():
+    from discord.ext import voice_recv
+
+    from src.voice._discord_voice_adapter import AudioSink
+
+    assert AudioSink is voice_recv.AudioSink
+
+
+def test_voice_data_is_voice_recv_voice_data():
+    from discord.ext import voice_recv
+
+    from src.voice._discord_voice_adapter import VoiceData
+
+    assert VoiceData is voice_recv.VoiceData
+
+
+# --- VoiceReceiveClient API surface used by the pipeline --------------------
+
+
+@pytest.mark.parametrize(
+    "method_name",
+    [
+        "connect",          # inherited from discord.VoiceClient
+        "disconnect",       # inherited from discord.VoiceClient
+        "listen",           # voice_recv extension
+        "stop_listening",   # voice_recv extension
+        "is_listening",     # voice_recv extension
+        "stop",             # voice_recv extension (stops play+listen)
+        "cleanup",          # voice_recv extension
+    ],
+)
+def test_voice_receive_client_has_method(method_name):
+    from src.voice._discord_voice_adapter import VoiceReceiveClient
+
+    attr = getattr(VoiceReceiveClient, method_name, None)
+    assert attr is not None, f"VoiceReceiveClient is missing `.{method_name}()`"
+    assert callable(attr), f"VoiceReceiveClient.{method_name} is not callable"
+
+
+def test_voice_receive_client_listen_accepts_sink_and_after():
+    """`.listen(sink, *, after=None)` is the canonical call shape."""
+    from src.voice._discord_voice_adapter import VoiceReceiveClient
+
+    sig = inspect.signature(VoiceReceiveClient.listen)
+    params = sig.parameters
+    assert "sink" in params, f"VoiceReceiveClient.listen missing `sink` param; got {list(params)}"
+    assert "after" in params, f"VoiceReceiveClient.listen missing `after` kwarg; got {list(params)}"
+
+
+def test_voice_receive_client_has_sink_property():
+    """`.sink` is read/write so we can swap sinks in place."""
+    from src.voice._discord_voice_adapter import VoiceReceiveClient
+
+    sink_attr = inspect.getattr_static(VoiceReceiveClient, "sink", None)
+    assert isinstance(sink_attr, property), "VoiceReceiveClient.sink must be a property"
+    assert sink_attr.fget is not None, "VoiceReceiveClient.sink property missing getter"
+    assert sink_attr.fset is not None, "VoiceReceiveClient.sink property missing setter"
+
+
+# --- AudioSink API surface --------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    "method_name",
+    [
+        "write",         # write(user, voice_data) — the hot path
+        "cleanup",
+        "wants_opus",    # bool: opus bytes vs decoded PCM
+    ],
+)
+def test_audio_sink_has_method(method_name):
+    from src.voice._discord_voice_adapter import AudioSink
+
+    attr = getattr(AudioSink, method_name, None)
+    assert attr is not None, f"AudioSink is missing `.{method_name}()`"
+    assert callable(attr), f"AudioSink.{method_name} is not callable"
+
+
+def test_audio_sink_write_signature():
+    """`.write(self, user, data)` — user is the speaker (Optional), data is VoiceData."""
+    from src.voice._discord_voice_adapter import AudioSink
+
+    sig = inspect.signature(AudioSink.write)
+    params = list(sig.parameters)
+    # self, user, data
+    assert len(params) >= 3, f"AudioSink.write expected (self, user, data), got {params}"
+
+
+# --- VoiceData attributes ---------------------------------------------------
+
+
+def test_voice_data_slots():
+    """VoiceData uses __slots__ for per-packet allocation. Pipeline reads these."""
+    from src.voice._discord_voice_adapter import VoiceData
+
+    assert hasattr(VoiceData, "__slots__"), "VoiceData lost __slots__ — perf regression risk"
+    slots = set(VoiceData.__slots__)
+    # Documented attributes the pipeline depends on.
+    assert "packet" in slots, f"VoiceData missing `packet` slot; got {slots}"
+    assert "source" in slots, f"VoiceData missing `source` slot (speaker user); got {slots}"
+    assert "pcm" in slots, f"VoiceData missing `pcm` slot (decoded audio); got {slots}"
+
+
+def test_voice_data_has_opus_property():
+    """`.opus` exposes the raw opus bytes from the underlying RTP packet."""
+    from src.voice._discord_voice_adapter import VoiceData
+
+    opus_attr = inspect.getattr_static(VoiceData, "opus", None)
+    assert isinstance(opus_attr, property), "VoiceData.opus must be a property"
+
+
+# --- Echo-core DAVE-decrypt fork guards -------------------------------------
+#
+# Two contract tests pinned by the DAVE receive-side decrypt patch.
+# See plan: /home/moltbot/.claude/plans/wiggly-exploring-glade.md
+#
+# These fail fast on either:
+#   1. An upstream voice-recv re-install wiping the fork's version marker
+#      (i.e. our patch is gone), OR
+#   2. A discord.py upgrade renaming the connection-level DAVE attrs the
+#      patch reads (`dave_session`, `dave_protocol_version`).
+
+
+def test_voice_recv_fork_version():
+    """Echo-core fork tag for the DAVE-decrypt patch.
+
+    Lane A bumps `voice_recv.__version__` to `'0.5.3a+echo.dave1'` (PEP 440
+    local segment). If this assertion fails after a vendor reinstall, the
+    fork patch has been lost — re-apply `_maybe_dave_decrypt` + the
+    `callback()` hook before deploying, or live voice will regress to the
+    `opus_decode: corrupted stream` error chain.
+    """
+    from discord.ext import voice_recv
+
+    assert voice_recv.__version__ == "0.5.3a+echo.dave1", (
+        f"voice_recv.__version__ is {voice_recv.__version__!r}; expected "
+        "'0.5.3a+echo.dave1'. The DAVE-decrypt fork patch has been "
+        "overwritten — re-apply before reinstalling the vendored package."
+    )
+
+
+def test_voice_connection_state_has_dave_attrs():
+    """`_maybe_dave_decrypt` reads `dave_session` and `dave_protocol_version`
+    off the discord.py `VoiceConnectionState`. If a future discord.py upgrade
+    renames either attr, fail loudly here rather than in a live voice call
+    (where the symptom is silent packet drops).
+    """
+    from discord import voice_state
+
+    src = inspect.getsource(voice_state.VoiceConnectionState)
+    assert "dave_session" in src, (
+        "discord.voice_state.VoiceConnectionState source no longer mentions "
+        "'dave_session' — discord.py may have renamed the attr. Update "
+        "vendor/discord-ext-voice-recv/.../reader.py::_maybe_dave_decrypt."
+    )
+    assert "dave_protocol_version" in src, (
+        "discord.voice_state.VoiceConnectionState source no longer mentions "
+        "'dave_protocol_version' — discord.py may have renamed the attr. "
+        "Update _maybe_dave_decrypt accordingly."
+    )

tests/test_voice_normalize.py

@@ -0,0 +1,137 @@
+"""Tests for src/voice/normalize.py — 35 Romanian cases.
+
+Categories:
+    markdown strip (5), numbers cardinals (6), decimals (4),
+    currency natural (8), symbols (4), abbreviations (4),
+    truncation boundary (2), edge cases empty / whitespace (2).
+
+Total: 35.
+"""
+import pytest
+
+from src.voice.normalize import (
+    expand_abbreviations,
+    expand_currency,
+    expand_numbers_ro,
+    expand_symbols,
+    normalize_for_tts,
+    strip_markdown,
+)
+
+
+# ============================================================
+# Markdown stripping (5)
+# ============================================================
+@pytest.mark.parametrize("text,expected", [
+    ("**bold text**", "bold text"),
+    ("*italic text*", "italic text"),
+    ("`code snippet`", "code snippet"),
+    ("[click here](https://example.com)", "click here"),
+    ("# Heading text", "Heading text"),
+])
+def test_strip_markdown(text, expected):
+    assert strip_markdown(text) == expected
+
+
+# ============================================================
+# Numbers cardinals (6)
+# ============================================================
+@pytest.mark.parametrize("text,expected", [
+    ("21", "douăzeci și unu"),
+    ("81", "optzeci și unu"),
+    ("100", "o sută"),
+    ("3", "trei"),
+    ("0", "zero"),
+    ("200", "două sute"),
+])
+def test_expand_numbers_cardinals(text, expected):
+    assert expand_numbers_ro(text) == expected
+
+
+# ============================================================
+# Decimals (4)
+# ============================================================
+@pytest.mark.parametrize("text,expected", [
+    ("3.14", "trei virgulă paisprezece"),
+    ("12.5", "doisprezece virgulă cinci"),
+    ("0.5", "zero virgulă cinci"),
+    ("99.99", "nouăzeci și nouă virgulă nouăzeci și nouă"),
+])
+def test_expand_numbers_decimals(text, expected):
+    assert expand_numbers_ro(text) == expected
+
+
+# ============================================================
+# Currency natural RO (8) — RON / USD / EUR / GBP mix
+# ============================================================
+@pytest.mark.parametrize("text,expected", [
+    ("12.50 RON", "doisprezece lei și cincizeci de bani"),
+    ("$25.99", "douăzeci și cinci de dolari și nouăzeci și nouă de cenți"),
+    ("€100.50", "o sută de euro și cincizeci de cenți"),
+    ("£200", "două sute de lire"),
+    ("100 RON", "o sută de lei"),
+    ("$1", "un dolar"),
+    ("€50", "cincizeci de euro"),
+    ("1 RON", "un leu"),
+])
+def test_expand_currency(text, expected):
+    assert expand_currency(text) == expected
+
+
+# ============================================================
+# Symbols (4)
+# ============================================================
+@pytest.mark.parametrize("text,expected", [
+    ("25%", "25 la sută"),
+    ("foo & bar", "foo și bar"),
+    ("Marius @ home", "Marius la home"),
+    ("30°", "30 grade"),
+])
+def test_expand_symbols(text, expected):
+    assert expand_symbols(text) == expected
+
+
+# ============================================================
+# Abbreviations (4)
+# ============================================================
+@pytest.mark.parametrize("text,expected", [
+    ("etc.", "etcetera"),
+    ("dl. Popescu", "domnul Popescu"),
+    ("dna. Ionescu", "doamna Ionescu"),
+    ("nr. 5", "numărul 5"),
+])
+def test_expand_abbreviations(text, expected):
+    assert expand_abbreviations(text) == expected
+
+
+# ============================================================
+# Truncation boundary (2)
+# ============================================================
+def test_truncate_exactly_200_words_unchanged():
+    """Exactly 200 simple word tokens — no truncation, no suffix."""
+    text = " ".join(["cuvant"] * 200)
+    out = normalize_for_tts(text)
+    assert "Restul l-am scris în chat." not in out
+    assert out.split() == ["cuvant"] * 200
+
+
+def test_truncate_over_200_words_appends_suffix():
+    """250 word tokens — keep first 200 then append the chat-deferral phrase."""
+    text = " ".join(["cuvant"] * 250)
+    out = normalize_for_tts(text)
+    assert out.endswith("Restul l-am scris în chat.")
+    words = out.split()
+    # First 200 are 'cuvant', followed by the 5-word suffix.
+    assert words[:200] == ["cuvant"] * 200
+    assert words[200:] == ["Restul", "l-am", "scris", "în", "chat."]
+
+
+# ============================================================
+# Edge cases (2)
+# ============================================================
+@pytest.mark.parametrize("text,expected", [
+    ("", ""),
+    ("   ", ""),
+])
+def test_normalize_edge_cases(text, expected):
+    assert normalize_for_tts(text) == expected

tests/test_voice_recv_dave.py

@@ -0,0 +1,302 @@
+"""DAVE receive-side decrypt tests for the vendored voice-recv fork.
+
+Exercises Lane A's patch on
+`vendor/discord-ext-voice-recv/discord/ext/voice_recv/reader.py`:
+
+* `_maybe_dave_decrypt(rtp_packet)` — DAVE E2E layer sandwiched between the
+  transport-layer decrypt and the routing into the opus decoder. No-op when
+  the room is non-DAVE, when davey isn't installed, or when the SSRC map
+  hasn't caught up to a new speaker yet.
+* `callback()` hook — feeds the DAVE-unwrapped plaintext into
+  `packet_router.feed_rtp()` on success, drops the packet on failure WITHOUT
+  killing the reader thread.
+
+The test fixtures mirror `tests/test_voice_session_cleanup.py:33-54`:
+  * Construct `AudioReader` via `AudioReader.__new__(AudioReader)` + manual
+    attr set so the reader thread is never started.
+  * `MagicMock` everything below the unit under test.
+
+`_HAS_DAVE` / `_MEDIA_TYPE_AUDIO` on the reader module are monkey-patched per
+test so the suite passes whether or not `davey` is importable in the venv.
+The assertions only become meaningful once Lane A's patch has landed and the
+package has been re-installed (`pip install -e vendor/discord-ext-voice-recv
+--force-reinstall`); the FILE itself is valid Python regardless.
+
+See plan: /home/moltbot/.claude/plans/wiggly-exploring-glade.md
+"""
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+
+from discord.ext.voice_recv.reader import AudioReader
+
+
+# Sentinel for `_MEDIA_TYPE_AUDIO`. Using a plain object() keeps the tests
+# independent of whether davey is importable — we just assert the value
+# flows through to `dave_session.decrypt()` unchanged.
+_FAKE_MEDIA_TYPE_AUDIO = object()
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def fake_dave_session():
+    sess = MagicMock(name="dave_session")
+    sess.ready = True
+    # Default: this user is NOT in passthrough — DAVE decrypt must run.
+    # Individual tests can override to True to exercise the passthrough path.
+    sess.can_passthrough = MagicMock(return_value=False)
+    sess.decrypt = MagicMock(return_value=b"plaintext_opus")
+    return sess
+
+
+@pytest.fixture
+def fake_connection(fake_dave_session):
+    conn = MagicMock(name="_connection")
+    conn.dave_protocol_version = 1
+    conn.dave_session = fake_dave_session
+    return conn
+
+
+@pytest.fixture
+def fake_voice_client(fake_connection):
+    vc = MagicMock(name="voice_client")
+    vc._connection = fake_connection
+    vc._ssrc_to_id = {12345: 999_000}
+    return vc
+
+
+@pytest.fixture
+def fake_rtp_packet():
+    pkt = MagicMock(name="rtp_packet")
+    pkt.ssrc = 12345
+    pkt.decrypted_data = b"ciphertext_after_transport_decrypt"
+    pkt.is_silence = MagicMock(return_value=False)
+    return pkt
+
+
+@pytest.fixture
+def reader(fake_voice_client):
+    """`AudioReader` instance with no reader thread spawned.
+
+    Same pattern used by `tests/test_voice_session_cleanup.py` for
+    `VoiceSession` — bypass `__init__` so we can drive the public surface
+    against pure mocks.
+    """
+    r = AudioReader.__new__(AudioReader)
+    r.voice_client = fake_voice_client
+    r.error = None
+    return r
+
+
+@pytest.fixture
+def dave_enabled(monkeypatch):
+    """Force the reader module's DAVE-availability flags ON.
+
+    Pins `_MEDIA_TYPE_AUDIO` to a known sentinel so the happy-path test can
+    assert exactly what gets passed to `dave_session.decrypt`. `raising=False`
+    keeps the monkeypatch valid even if Lane A's patch hasn't landed yet —
+    the tests will still fail (no `_maybe_dave_decrypt` attr), just for the
+    right reason.
+    """
+    import discord.ext.voice_recv.reader as reader_mod
+
+    monkeypatch.setattr(reader_mod, "_HAS_DAVE", True, raising=False)
+    monkeypatch.setattr(
+        reader_mod, "_MEDIA_TYPE_AUDIO", _FAKE_MEDIA_TYPE_AUDIO, raising=False
+    )
+    return reader_mod
+
+
+# ---------------------------------------------------------------------------
+# Unit tests: `_maybe_dave_decrypt`
+# ---------------------------------------------------------------------------
+
+
+class TestMaybeDaveDecrypt:
+    """Seven unit tests on the DAVE-decrypt gate.
+
+    The gate mirrors `voice_client.can_encrypt` in discord.py 2.7.1 exactly
+    (`voice_state.py:272-273`). Bypass semantics on every "DAVE inactive"
+    branch let non-DAVE rooms and davey-less environments keep working.
+    """
+
+    def test_protocol_version_zero_bypasses_decrypt(
+        self, dave_enabled, reader, fake_connection, fake_dave_session, fake_rtp_packet,
+    ):
+        """`dave_protocol_version == 0` → return the transport-decrypted
+        payload unchanged; never touch `dave_session.decrypt`."""
+        fake_connection.dave_protocol_version = 0
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result is fake_rtp_packet.decrypted_data
+        fake_dave_session.decrypt.assert_not_called()
+
+    def test_dave_session_none_bypasses_decrypt(
+        self, dave_enabled, reader, fake_connection, fake_rtp_packet,
+    ):
+        """`dave_session is None` → bypass. Pre-MLS-handshake state."""
+        fake_connection.dave_session = None
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result is fake_rtp_packet.decrypted_data
+
+    def test_dave_session_not_ready_bypasses_decrypt(
+        self, dave_enabled, reader, fake_dave_session, fake_rtp_packet,
+    ):
+        """`dave_session.ready is False` → bypass. Pre-MLS-epoch-1 packets
+        are transport-only on the wire."""
+        fake_dave_session.ready = False
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result is fake_rtp_packet.decrypted_data
+        fake_dave_session.decrypt.assert_not_called()
+
+    def test_unknown_ssrc_returns_none(
+        self, dave_enabled, reader, fake_voice_client, fake_dave_session, fake_rtp_packet,
+    ):
+        """SSRC not in `_ssrc_to_id` → drop (return None).
+
+        Accepted regression: davey requires per-user keys; when SPEAKING
+        events race behind the first audio packet, 1-5 packets per new
+        speaker per session are dropped. See plan §Edge cases.
+        """
+        fake_voice_client._ssrc_to_id.clear()
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result is None
+        fake_dave_session.decrypt.assert_not_called()
+
+    def test_happy_path_invokes_decrypt_and_returns_plaintext(
+        self, dave_enabled, reader, fake_dave_session, fake_rtp_packet,
+    ):
+        """Full DAVE-active path: `decrypt(user_id, MediaType.audio, ciphertext)`
+        called exactly once with the expected args; method returns the
+        davey plaintext bytes verbatim."""
+        ciphertext = fake_rtp_packet.decrypted_data
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result == b"plaintext_opus"
+        fake_dave_session.decrypt.assert_called_once_with(
+            999_000, _FAKE_MEDIA_TYPE_AUDIO, ciphertext,
+        )
+
+    def test_decrypt_raises_returns_none_no_crash(
+        self, dave_enabled, reader, fake_dave_session, fake_rtp_packet,
+    ):
+        """davey.decrypt raising → drop the packet, don't propagate, and
+        leave `reader.error` untouched so the reader thread stays alive.
+
+        MLS epoch transitions can produce transient decrypt failures —
+        bumping `reader.error` would call `self.stop()` and kill the whole
+        receive pipeline."""
+        fake_dave_session.decrypt.side_effect = RuntimeError(
+            "simulated MLS epoch transition fail"
+        )
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result is None
+        assert reader.error is None
+
+    def test_has_dave_false_bypasses_even_with_session_present(
+        self, monkeypatch, reader, fake_dave_session, fake_rtp_packet,
+    ):
+        """`_HAS_DAVE = False` → bypass everything, even if a real session
+        somehow showed up on the connection. Defensive shim that keeps the
+        tests (and any davey-less deploys) green."""
+        import discord.ext.voice_recv.reader as reader_mod
+
+        monkeypatch.setattr(reader_mod, "_HAS_DAVE", False, raising=False)
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result is fake_rtp_packet.decrypted_data
+        fake_dave_session.decrypt.assert_not_called()
+
+    def test_can_passthrough_true_returns_payload_without_decrypt(
+        self, dave_enabled, reader, fake_dave_session, fake_rtp_packet,
+    ):
+        """`can_passthrough(user_id) == True` → return the transport-decrypted
+        payload as-is; never call `decrypt`. Mirrors Discord's protocol where
+        a passthrough-mode peer sends non-DAVE-wrapped packets that the
+        receiver must accept verbatim."""
+        fake_dave_session.can_passthrough = MagicMock(return_value=True)
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result is fake_rtp_packet.decrypted_data
+        fake_dave_session.can_passthrough.assert_called_once_with(999_000)
+        fake_dave_session.decrypt.assert_not_called()
+
+    def test_can_passthrough_raises_falls_through_to_decrypt(
+        self, dave_enabled, reader, fake_dave_session, fake_rtp_packet,
+    ):
+        """`can_passthrough` raising → swallow the error and try `decrypt`.
+        Defensive: an older davey build or transient internal state shouldn't
+        break the receive pipeline."""
+        fake_dave_session.can_passthrough = MagicMock(
+            side_effect=RuntimeError("simulated davey internal error")
+        )
+        result = reader._maybe_dave_decrypt(fake_rtp_packet)
+        assert result == b"plaintext_opus"
+        fake_dave_session.decrypt.assert_called_once()
+
+
+# ---------------------------------------------------------------------------
+# Integration tests: `callback()` exercises the DAVE hook
+# ---------------------------------------------------------------------------
+
+
+class TestCallbackIntegration:
+    """Two integration tests for the hook Lane A inserts between transport
+    decrypt (reader.py:141) and the post-decrypt routing (reader.py:159).
+
+    Strategy: stub the transport-decrypt and RTP parsing path so `callback()`
+    reaches the hook, then mock `_maybe_dave_decrypt` directly on the reader
+    instance. The assertion focuses on `feed_rtp` being called (test 8) vs.
+    not called (test 9). The transport path correctness is covered by
+    voice-recv's own upstream tests.
+    """
+
+    @staticmethod
+    def _wire_callback(reader, monkeypatch, fake_rtp_packet):
+        import discord.ext.voice_recv.reader as reader_mod
+
+        # Redirect rtp parsing — we want an RTP path (not RTCP) so the hook fires.
+        monkeypatch.setattr(reader_mod.rtp, "is_rtcp", lambda data: False)
+        monkeypatch.setattr(reader_mod.rtp, "decode_rtp", lambda data: fake_rtp_packet)
+
+        # Stub the instance attrs `callback()` touches besides the hook.
+        reader.decryptor = MagicMock(name="decryptor")
+        reader.decryptor.decrypt_rtp = MagicMock(return_value=b"ciphertext")
+        reader.packet_router = MagicMock(name="packet_router")
+        reader.packet_router.feed_rtp = MagicMock()
+        reader.speaking_timer = MagicMock(name="speaking_timer")
+        reader.sink = MagicMock(name="sink")
+
+    def test_callback_feeds_when_dave_returns_bytes(
+        self, monkeypatch, reader, fake_rtp_packet,
+    ):
+        """Hook returns plaintext → `feed_rtp` called once with the
+        rtp_packet whose `decrypted_data` is now the post-DAVE plaintext."""
+        self._wire_callback(reader, monkeypatch, fake_rtp_packet)
+        plaintext = b"dave_unwrapped_opus_payload"
+        reader._maybe_dave_decrypt = MagicMock(return_value=plaintext)
+
+        reader.callback(b"raw_packet_bytes")
+
+        reader._maybe_dave_decrypt.assert_called_once_with(fake_rtp_packet)
+        assert reader.packet_router.feed_rtp.call_count == 1
+        called_with = reader.packet_router.feed_rtp.call_args[0][0]
+        assert called_with is fake_rtp_packet
+        assert fake_rtp_packet.decrypted_data == plaintext
+        assert reader.error is None
+
+    def test_callback_drops_when_dave_returns_none(
+        self, monkeypatch, reader, fake_rtp_packet,
+    ):
+        """Hook returns None → `feed_rtp` NOT called, no exception propagated,
+        `reader.error` stays None (reader thread survives the drop)."""
+        self._wire_callback(reader, monkeypatch, fake_rtp_packet)
+        reader._maybe_dave_decrypt = MagicMock(return_value=None)
+
+        reader.callback(b"raw_packet_bytes")
+
+        reader._maybe_dave_decrypt.assert_called_once_with(fake_rtp_packet)
+        reader.packet_router.feed_rtp.assert_not_called()
+        assert reader.error is None

tests/test_voice_session_cleanup.py

@@ -0,0 +1,319 @@
+"""Cleanup-path tests for ``src/voice/pipeline.py::VoiceSession``.
+
+Pins the centralized ``cleanup()`` contract from the voice plan
+(Engineering decision #5): every one of the FIVE exit paths must drain
+state cleanly and idempotently — lock released, JSONL flushed or
+discarded, presence cleared, ``voice_client.cleanup()`` invoked,
+``ttsq.stop()`` invoked, and a second call to ``cleanup()`` MUST be a
+no-op (side effects happen exactly once).
+
+The 5 paths under test:
+    1. ``test_cleanup_on_voice_leave``           — explicit ``/voice leave``
+    2. ``test_cleanup_on_disconnect``            — Discord-level disconnect
+    3. ``test_cleanup_on_crash``                 — exception via ``__exit__``
+    4. ``test_cleanup_on_auto_leave``            — 5-min inactivity timer
+    5. ``test_cleanup_on_user_leaves_channel``   — user leaves voice channel
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from src.voice.pipeline import VoiceSession
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def mock_bot():
+    bot = MagicMock(name="bot")
+    bot.user = MagicMock()
+    bot.user.id = 999_999
+    bot.change_presence = AsyncMock(name="change_presence")
+    bot.get_user = MagicMock(return_value=None)
+    return bot
+
+
+@pytest.fixture
+def mock_voice_client():
+    vc = MagicMock(name="voice_client")
+    vc.cleanup = MagicMock(name="vc_cleanup")
+    return vc
+
+
+@pytest.fixture
+def mock_ttsq():
+    ttsq = MagicMock(name="ttsq")
+    ttsq.stop = MagicMock(name="ttsq_stop")
+    return ttsq
+
+
+@pytest.fixture
+def mock_text_channel():
+    tc = MagicMock(name="text_channel")
+    tc.send = AsyncMock(name="text_send")
+    return tc
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _make_session(
+    tmp_path: Path,
+    mock_bot,
+    mock_voice_client,
+    mock_ttsq,
+    mock_text_channel,
+    *,
+    record_enabled: bool = True,
+) -> VoiceSession:
+    jsonl = tmp_path / ("transcripts.jsonl" if record_enabled else "noop.jsonl")
+    return VoiceSession(
+        channel_id=1001,
+        guild_id=42,
+        text_channel=mock_text_channel,
+        voice_client=mock_voice_client,
+        bot=mock_bot,
+        ttsq=mock_ttsq,
+        whitelist={1234},
+        record_enabled=record_enabled,
+        mirror_enabled=True,
+        transcripts_jsonl_path=jsonl,
+        loop=None,
+        router_route_message=MagicMock(name="route_message"),
+    )
+
+
+def _assert_clean_post_cleanup(
+    session: VoiceSession,
+    voice_client,
+    ttsq,
+    bot,
+    jsonl_path: Path,
+    record_enabled: bool,
+) -> None:
+    """Assertions shared across all five cleanup-path tests."""
+    # 1. Lock released — non-blocking acquire from this thread returns True.
+    acquired = session._lock.acquire(blocking=False)
+    assert acquired, "session._lock must be released after cleanup()"
+    session._lock.release()
+
+    # 2. voice_client.cleanup() called exactly once.
+    assert voice_client.cleanup.call_count == 1, (
+        f"voice_client.cleanup() called {voice_client.cleanup.call_count}x, "
+        f"expected 1"
+    )
+
+    # 3. ttsq.stop() called exactly once.
+    assert ttsq.stop.call_count == 1, (
+        f"ttsq.stop() called {ttsq.stop.call_count}x, expected 1"
+    )
+
+    # 4. bot.change_presence(activity=None) called at least once with that kwarg.
+    assert bot.change_presence.call_count >= 1, (
+        "bot.change_presence was never called — presence not restored"
+    )
+    bot.change_presence.assert_called_with(activity=None)
+
+    # 5. JSONL flushed (record=on) OR absent (record=off).
+    if record_enabled:
+        assert jsonl_path.exists(), (
+            "record=on: JSONL file must exist (was created by __enter__ and "
+            "left in place by cleanup so transcript can be persisted)"
+        )
+    else:
+        # record=off: cleanup unlinks the file if it ever existed.
+        assert not jsonl_path.exists() or jsonl_path.stat().st_size == 0
+
+
+# ---------------------------------------------------------------------------
+# Scenario 1 — explicit /voice leave
+# ---------------------------------------------------------------------------
+
+
+class TestCleanupOnVoiceLeave:
+    def test_cleanup_on_voice_leave(
+        self, tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+    ):
+        session = _make_session(
+            tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+            record_enabled=True,
+        )
+        jsonl_path = session.transcripts_jsonl_path
+
+        with session:
+            # Simulate one transcript line.
+            session._jsonl_fh.write(json.dumps({"text": "salut"}) + "\n")
+            session.cleanup("voice_leave")
+            assert session._cleaned_up is True
+
+        # __exit__ called cleanup("exit") — must be a no-op the second time.
+        _assert_clean_post_cleanup(
+            session, mock_voice_client, mock_ttsq, mock_bot,
+            jsonl_path, record_enabled=True,
+        )
+
+        # Idempotency: a third explicit call still doesn't bump counts.
+        session.cleanup("redundant")
+        assert mock_voice_client.cleanup.call_count == 1
+        assert mock_ttsq.stop.call_count == 1
+
+
+# ---------------------------------------------------------------------------
+# Scenario 2 — Discord-level voice disconnect
+# ---------------------------------------------------------------------------
+
+
+class TestCleanupOnDisconnect:
+    def test_cleanup_on_disconnect(
+        self, tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+    ):
+        session = _make_session(
+            tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+            record_enabled=False,
+        )
+        jsonl_path = session.transcripts_jsonl_path
+
+        session.__enter__()
+        # Network drop arrives outside the with-block.
+        session.cleanup("disconnect")
+        _assert_clean_post_cleanup(
+            session, mock_voice_client, mock_ttsq, mock_bot,
+            jsonl_path, record_enabled=False,
+        )
+
+        # Idempotency.
+        session.cleanup("disconnect-again")
+        assert mock_voice_client.cleanup.call_count == 1
+        assert mock_ttsq.stop.call_count == 1
+
+
+# ---------------------------------------------------------------------------
+# Scenario 3 — crash / exception via __exit__
+# ---------------------------------------------------------------------------
+
+
+class TestCleanupOnCrash:
+    def test_cleanup_on_crash(
+        self, tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+    ):
+        session = _make_session(
+            tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+            record_enabled=True,
+        )
+        jsonl_path = session.transcripts_jsonl_path
+
+        with pytest.raises(RuntimeError, match="simulated crash"):
+            with session:
+                # Pipeline raises mid-call.
+                raise RuntimeError("simulated crash")
+
+        # __exit__ must have driven cleanup — every side effect happened once.
+        _assert_clean_post_cleanup(
+            session, mock_voice_client, mock_ttsq, mock_bot,
+            jsonl_path, record_enabled=True,
+        )
+
+        # Idempotency: explicit follow-up call (e.g. an outer error handler
+        # also tries to cleanup) MUST be a no-op.
+        session.cleanup("post-crash")
+        assert mock_voice_client.cleanup.call_count == 1
+        assert mock_ttsq.stop.call_count == 1
+
+
+# ---------------------------------------------------------------------------
+# Scenario 4 — auto-leave timer fires after 5 min inactivity
+# ---------------------------------------------------------------------------
+
+
+class TestCleanupOnAutoLeave:
+    def test_cleanup_on_auto_leave(
+        self, tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+    ):
+        session = _make_session(
+            tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+            record_enabled=True,
+        )
+        jsonl_path = session.transcripts_jsonl_path
+
+        session.__enter__()
+        # The auto-leave timer trips outside the with-block.
+        session.cleanup("auto_leave")
+
+        _assert_clean_post_cleanup(
+            session, mock_voice_client, mock_ttsq, mock_bot,
+            jsonl_path, record_enabled=True,
+        )
+
+        # Idempotency.
+        session.cleanup("auto_leave_redundant")
+        assert mock_voice_client.cleanup.call_count == 1
+        assert mock_ttsq.stop.call_count == 1
+
+
+# ---------------------------------------------------------------------------
+# Scenario 5 — user leaves voice channel themselves
+# ---------------------------------------------------------------------------
+
+
+class TestCleanupOnUserLeaves:
+    def test_cleanup_on_user_leaves_channel(
+        self, tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+    ):
+        session = _make_session(
+            tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+            record_enabled=False,
+        )
+        jsonl_path = session.transcripts_jsonl_path
+
+        session.__enter__()
+        # voice_state_update event handler invokes cleanup directly.
+        session.cleanup("user_left_channel")
+
+        _assert_clean_post_cleanup(
+            session, mock_voice_client, mock_ttsq, mock_bot,
+            jsonl_path, record_enabled=False,
+        )
+
+        # Idempotency.
+        session.cleanup("user_left_again")
+        assert mock_voice_client.cleanup.call_count == 1
+        assert mock_ttsq.stop.call_count == 1
+
+
+# ---------------------------------------------------------------------------
+# Cross-cutting: failures inside cleanup don't propagate
+# ---------------------------------------------------------------------------
+
+
+class TestCleanupRobustness:
+    def test_cleanup_swallows_voice_client_errors(
+        self, tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+    ):
+        """If voice_client.cleanup() raises, ttsq.stop() must still run and
+        the lock must still release — otherwise a broken Discord state would
+        deadlock the channel forever."""
+        mock_voice_client.cleanup.side_effect = RuntimeError("vc died")
+
+        session = _make_session(
+            tmp_path, mock_bot, mock_voice_client, mock_ttsq, mock_text_channel,
+            record_enabled=False,
+        )
+
+        with session:
+            session.cleanup("voice_leave")
+
+        # ttsq.stop still ran exactly once.
+        assert mock_ttsq.stop.call_count == 1
+        # Lock released.
+        acquired = session._lock.acquire(blocking=False)
+        assert acquired, "lock must release even when voice_client.cleanup raises"
+        session._lock.release()

tools/tts.py

@@ -23,6 +23,24 @@ VOICES = {"M1", "M2", "M3", "M4", "M5", "F1", "F2", "F3", "F4", "F5"}
 DEFAULT_VOICE = "M2"
 DEFAULT_LANG = "ro"
 
+# Punctuation Supertonic synthesis rejects with HTTP 500 (Romanian curly quotes,
+# smart dashes, ellipsis, angle quotes). Mapped to ASCII so a stray „foo" in
+# any caller's text doesn't kill the whole request.
+_TTS_PUNCT_MAP = {
+    '„': '"', '“': '"', '”': '"',
+    '‘': "'", '’': "'", '‚': "'",
+    '«': '"', '»': '"',
+    '–': '-', '—': '-',
+    '…': '...',
+}
+
+
+def sanitize_for_supertonic(text: str) -> str:
+    """Replace Unicode punctuation Supertonic rejects with ASCII equivalents."""
+    for src, dst in _TTS_PUNCT_MAP.items():
+        text = text.replace(src, dst)
+    return text
+
 
 def synthesize(text: str, voice: str = DEFAULT_VOICE, lang: str = DEFAULT_LANG) -> dict:
     """Call Supertonic server and save audio to a temp WAV file.
@@ -34,6 +52,8 @@ def synthesize(text: str, voice: str = DEFAULT_VOICE, lang: str = DEFAULT_LANG)
     if not text or not text.strip():
         return {"ok": False, "error": "Text gol."}
 
+    text = sanitize_for_supertonic(text)
+
     voice = voice.upper()
     if voice not in VOICES:
         voice = DEFAULT_VOICE

vendor/discord-ext-voice-recv/VENDOR_INFO.md

@@ -1,22 +1,76 @@
 # Vendored: discord-ext-voice-recv
 
 **Upstream:** https://github.com/imayhaveborkedit/discord-ext-voice-recv
-**Pinned commit:** `ac04ea7b0941112e83767cf1c1469b408fa06748` (bump version 0.5.3a)
+**Pinned commit:** `ac04ea7b0941112e83767cf1c1469b408fa06748` (bump version 0.5.3a, master HEAD Jun 2025)
 **Vendored at:** 2026-05-27
+**Echo Core fork version:** `0.5.3a+echo.dave1` (PEP 440 local segment)
 **Reason:** Discord voice protocol is fragile, upstream is hobby fork. Adapter
 layer in `src/voice/_discord_voice_adapter.py` isolates upstream churn — if this
 package breaks, swap to py-cord by rewriting only that file.
 
-## Update procedure
+## Echo Core patch: `+echo.dave1` (DAVE E2E receive-side decrypt)
+
+### Why
+
+Discord enforces DAVE (E2E media encryption) on voice gateway `v=8` whenever the
+bot advertises `max_dave_protocol_version > 0` in IDENTIFY. discord.py 2.7.1 (the
+version Echo Core pins) does so unconditionally — Discord then closes the WS
+with code **4017** if the bot opts out by sending `max_dave_protocol_version=0`.
+DAVE is **mandatory**.
+
+Audio received from a DAVE-active room is **dual-wrapped**: transport layer
+(`aead_xchacha20_poly1305_rtpsize`) + DAVE E2E. Upstream voice-recv decrypts
+only the transport layer, then hands DAVE ciphertext to libopus, which raises
+`OpusError: corrupted stream` on every packet.
+
+### Patch shape
+
+~30 lines, all in `discord/ext/voice_recv/reader.py`:
+
+1. Module-level optional `davey` import (no-op when missing).
+2. `AudioReader._maybe_dave_decrypt(rtp_packet) -> Optional[bytes]` — gate logic
+   mirrors discord.py 2.7.1 send-side `can_encrypt` exactly. Returns the
+   DAVE-unwrapped payload, the original payload (DAVE inactive), or `None` to
+   drop the packet (unknown SSRC, decrypt failure).
+3. 4-line hook in `callback()` between transport-decrypt and `feed_rtp`:
+   overwrites `rtp_packet.decrypted_data` in place, or returns early to drop.
+
+The post-decrypt `is_silence()` check (formerly at reader.py:172) still works
+because we overwrite `decrypted_data` in place — silence frames produced by
+davey reach the existing check unchanged.
+
+### Dependency
+
+`davey==0.1.5` — matches discord.py 2.7.1 expectation. Pin in
+`echo-core/requirements.txt`. The import is optional at module level so tests
+and non-DAVE environments still run; the gate degrades to a bypass.
+
+### Re-sync strategy
+
+When upstream voice-recv adds DAVE support natively:
+
+1. Drop the three patch hunks in `reader.py` (davey import block,
+   `_maybe_dave_decrypt` method, hook in `callback()`).
+2. Revert `__version__` to upstream value in `__init__.py`.
+3. Update `Pinned commit` below.
+4. Run `pytest tests/test_voice_recv_dave.py tests/test_voice_adapter_contract.py`.
+
+The contract test `test_voice_recv_fork_version` asserts `__version__ ==
+'0.5.3a+echo.dave1'` and will fail fast on any accidental wipe during a careless
+upstream sync — forcing a conscious decision to either re-port or drop the
+patch.
+
+## Update procedure (vanilla upstream sync)
 
 ```bash
 cd vendor/discord-ext-voice-recv
 git fetch origin master
 git log HEAD..origin/master --oneline  # review what changed
 git checkout <new-commit>
+# RE-APPLY the +echo.dave1 patch if upstream still lacks DAVE
 cd ../..
 source .venv/bin/activate && pip install -e vendor/discord-ext-voice-recv --force-reinstall
-pytest tests/test_voice_adapter_contract.py -v  # MUST PASS — contract guard
+pytest tests/test_voice_adapter_contract.py tests/test_voice_recv_dave.py -v  # MUST PASS — contract + DAVE guards
 ```
 
 Update this file's `Pinned commit` after a successful upgrade.

vendor/discord-ext-voice-recv/discord/ext/voice_recv/__init__.py

@@ -17,4 +17,4 @@ __title__ = 'discord.ext.voice_recv'
 __author__ = 'Imayhaveborkedit'
 __license__ = 'MIT'
 __copyright__ = 'Copyright 2021-present Imayhaveborkedit'
-__version__ = '0.5.3a'
+__version__ = '0.5.3a+echo.dave1'

vendor/discord-ext-voice-recv/discord/ext/voice_recv/reader.py

@@ -19,6 +19,15 @@ try:
 except ImportError as e:
     raise RuntimeError("pynacl is required") from e
 
+# Echo Core +echo.dave1 patch: DAVE E2E receive-side decrypt. See VENDOR_INFO.md.
+try:
+    import davey
+    _MEDIA_TYPE_AUDIO = davey.MediaType.audio
+    _HAS_DAVE = True
+except ImportError:
+    _MEDIA_TYPE_AUDIO = None
+    _HAS_DAVE = False
+
 if TYPE_CHECKING:
     from typing import Optional, Callable, Any, Dict, Literal, Union
 
@@ -133,12 +142,63 @@ class AudioReader:
     def _is_ip_discovery_packet(self, data: bytes) -> bool:
         return len(data) == 74 and data[1] == 0x02
 
+    def _maybe_dave_decrypt(self, rtp_packet) -> Optional[bytes]:
+        """DAVE E2E layer applied after transport decrypt.
+
+        Returns the (possibly DAVE-unwrapped) opus payload, or None to drop the
+        packet. No-op when DAVE is inactive — non-DAVE rooms and environments
+        without `davey` installed pass through unchanged.
+
+        NOTE: `is_silence()` is NOT checked here. In a DAVE-active room the
+        transport-decrypted payload is ciphertext, so `is_silence()` (which
+        compares to plaintext OPUS_SILENCE ``b'\\xf8\\xff\\xfe'``) never matches.
+        Silence frames are handled either by davey.decrypt returning plaintext
+        silence (then caught at the existing post-decrypt silence check on
+        ``decrypted_data``), or dropped via the decrypt-raises path. The
+        existing post-decrypt silence check continues to work because we
+        overwrite ``decrypted_data`` in place.
+        """
+        if not _HAS_DAVE:
+            return rtp_packet.decrypted_data
+        conn = self.voice_client._connection
+        if getattr(conn, 'dave_protocol_version', 0) == 0:
+            return rtp_packet.decrypted_data
+        dave = getattr(conn, 'dave_session', None)
+        if dave is None or not dave.ready:
+            return rtp_packet.decrypted_data
+        user_id = self.voice_client._ssrc_to_id.get(rtp_packet.ssrc)
+        if user_id is None:
+            # ACCEPTED REGRESSION: davey requires per-user key. When SPEAKING
+            # event races behind the first audio packet, we drop 1-5 packets
+            # (~40-200ms) per new speaker per session.
+            return None
+        # can_passthrough(user_id) mirrors Discord's protocol: when this user's
+        # decryptor is in passthrough mode, packets are not DAVE-wrapped and
+        # must be returned as-is. Otherwise davey.decrypt unwraps DAVE E2E.
+        try:
+            if dave.can_passthrough(user_id):
+                return rtp_packet.decrypted_data
+        except Exception as e:
+            log.debug("can_passthrough check failed for ssrc=%s user=%s: %s: %s",
+                      rtp_packet.ssrc, user_id, type(e).__name__, e)
+        try:
+            return dave.decrypt(user_id, _MEDIA_TYPE_AUDIO, rtp_packet.decrypted_data)
+        except Exception as e:
+            log.debug("DAVE decrypt failed for ssrc=%s user=%s: %s: %s",
+                      rtp_packet.ssrc, user_id, type(e).__name__, e)
+            return None
+
     def callback(self, packet_data: bytes) -> None:
         packet = rtp_packet = rtcp_packet = None
         try:
             if not rtp.is_rtcp(packet_data):
                 packet = rtp_packet = rtp.decode_rtp(packet_data)
                 packet.decrypted_data = self.decryptor.decrypt_rtp(packet)
+                # Echo Core +echo.dave1: DAVE E2E layer (no-op when inactive).
+                dave_payload = self._maybe_dave_decrypt(rtp_packet)
+                if dave_payload is None:
+                    return  # drop packet, do not feed_rtp; reader thread stays alive
+                rtp_packet.decrypted_data = dave_payload
             else:
                 packet = rtcp_packet = rtp.decode_rtcp(self.decryptor.decrypt_rtcp(packet_data))