Pin remaining unbounded deps, drop stale JS model-list fallback, add smoke test

pyannote.audio was still `>=3.1` — the same unbounded-constraint pattern
that silently broke qwen3 (transformers>=4.45.0 resolving to a version
without qwen3_asr support on a later rebuild). Confirmed it had already
drifted once this session (4.0.4 -> 4.0.7). Pinned to the exact version
the current diarization code (token=, .speaker_diarization) is verified
against.

asr.js's config-fetch failure path fabricated a hardcoded backend/model
list that could silently drift from the server's real one (this is
exactly how the wrong Qwen3-ASR-2B/8B model IDs stuck around). Replaced it
with a visible error instead of a second source of truth.

Added scripts/smoke_test.sh: hits every backend's /transcribe or
/synthesize directly with a synthetic clip. Every regression found this
session (wrong model IDs, pyannote API drift, cuDNN path conflict) would
have shown up here immediately instead of waiting for a user to hit it.

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

app/ui/assets/js/asr.js

@@ -7,16 +7,13 @@ async function loadConfig() {
     const r = await fetch('/asr/config');
     SERVER_CONFIG = await r.json();
   } catch (e) {
-    SERVER_CONFIG = {
+    // No hardcoded fallback list here on purpose: the server's /asr/config is
-      default_backend: 'faster-whisper',
+    // the single source of truth for backend/model names. A duplicated guess
-      default_model: 'large-v3',
+    // here would silently drift out of sync with it (this has already bitten
-      default_language: 'ko',
+    // us once — see qwen3 model IDs). Surface the failure instead.
-      backends: {
+    SERVER_CONFIG = null;
-        'faster-whisper': { models: ['tiny','base','small','medium','large-v3','large-v2','turbo'] },
+    const el = document.getElementById('file-status');
-        'qwen3': { models: ['Qwen/Qwen3-ASR-0.6B-hf','Qwen/Qwen3-ASR-1.7B-hf'] },
+    if (el) el.textContent = '설정을 불러오지 못했습니다. 페이지를 새로고침해주세요.';
-        'vibevoice': { models: ['microsoft/VibeVoice-ASR'] },
-      },
-    };
   }
   initAsrUI();
 }

envs/faster_whisper.txt

@@ -2,5 +2,5 @@ fastapi==0.115.0
 uvicorn[standard]==0.30.6
 python-multipart==0.0.9
 faster-whisper==1.1.1
-pyannote.audio>=3.1
+pyannote.audio==4.0.7
 numpy

envs/qwen3.txt

@@ -2,7 +2,7 @@
 fastapi==0.115.0
 uvicorn[standard]==0.30.6
 python-multipart==0.0.9
-transformers>=4.45.0
+transformers==5.14.1
 accelerate>=0.30.0
 librosa>=0.10.0
 soundfile>=0.12.0

scripts/smoke_test.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Minimal post-deploy smoke test. Hits every ASR/TTS worker's own endpoint
+# directly (bypassing the gateway's OIDC auth) with a synthetic test clip, to
+# catch startup/dependency/API-compat regressions before a user has to find
+# them (qwen3's wrong model IDs, pyannote's use_auth_token/token rename, and
+# the cuDNN LD_LIBRARY_PATH conflict all would have shown up here as an
+# immediate FAIL instead of a silent break discovered later).
+#
+# Run this after every `bash build.sh` + redeploy.
+set -uo pipefail
+
+CONTAINER="${SPEECH_CONTAINER:-speech}"
+TMP_WAV="/tmp/smoke_test_tone.wav"
+FAIL=0
+
+pass() { printf '  OK  %s\n' "$1"; }
+fail() { printf 'FAIL  %s: %s\n' "$1" "$2"; FAIL=1; }
+
+echo "=== generating synthetic test clip ==="
+podman exec "$CONTAINER" ffmpeg -y -f lavfi -i "sine=frequency=440:duration=3" -ar 16000 -ac 1 "$TMP_WAV" >/dev/null 2>&1
+
+echo "=== faster-whisper (8001) ==="
+resp=$(podman exec "$CONTAINER" curl -s -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8001/transcribe \
+  -F "file=@${TMP_WAV};type=audio/wav" -F "model=tiny" -F "language=ko")
+[ "$resp" = "200" ] && pass "faster-whisper" || fail "faster-whisper" "HTTP $resp"
+
+echo "=== qwen3 (8004) ==="
+resp=$(podman exec "$CONTAINER" curl -s -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8004/transcribe \
+  -F "file=@${TMP_WAV};type=audio/wav" -F "model=Qwen/Qwen3-ASR-0.6B-hf" -F "language=ko")
+[ "$resp" = "200" ] && pass "qwen3" || fail "qwen3" "HTTP $resp"
+
+echo "=== vibevoice (8006) ==="
+resp=$(podman exec "$CONTAINER" curl -s -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8006/transcribe \
+  -F "file=@${TMP_WAV};type=audio/wav" -F "model=microsoft/VibeVoice-ASR")
+[ "$resp" = "200" ] && pass "vibevoice" || fail "vibevoice" "HTTP $resp"
+
+echo "=== xtts (8005) ==="
+resp=$(podman exec "$CONTAINER" curl -s -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8005/synthesize \
+  -F "text=스모크 테스트입니다." -F "language=ko" -F "speaker_wav=${TMP_WAV}")
+[ "$resp" = "200" ] && pass "xtts" || fail "xtts" "HTTP $resp"
+
+podman exec "$CONTAINER" rm -f "$TMP_WAV"
+
+echo
+if [ "$FAIL" = "0" ]; then
+  echo "all backends OK"
+else
+  echo "one or more backends FAILED — check: podman logs $CONTAINER"
+fi
+exit "$FAIL"