chore(audio): generate Vietnamese clips for both voices

- Run scripts/generate-audio.py to produce 184 MP3 clips (92 each
  for vi-VN-HoaiMyNeural and vi-VN-NamMinhNeural), ~2.2 MB total.
- Cap edge-tts concurrency at 4 with 4-attempt exponential retry on
  NoAudioReceived — earlier all-at-once gather() hit the upstream
  rate limit and bailed mid-voice.
- .gitignore: add .venv/ + __pycache__/ for the local generator venv.

web/.gitignore

@@ -5,3 +5,5 @@ build/
 .env.local
 .env*.local
 repomix-output.xml
+.venv/
+__pycache__/

web/scripts/generate-audio.py

@@ -57,10 +57,37 @@ def display_label(short_name: str, gender: str) -> str:
     return f"{given} ({gender_vi})"
 
 
+# Microsoft's TTS endpoint rate-limits aggressive concurrency and
+# occasionally returns empty streams. Cap parallelism + retry transient
+# failures so one voice run doesn't bail halfway through.
+CONCURRENCY = 4
+MAX_RETRIES = 4
+RETRY_BACKOFF_SEC = 1.5
+
+_sem: asyncio.Semaphore | None = None
+
+
+def _semaphore() -> asyncio.Semaphore:
+    global _sem
+    if _sem is None:
+        _sem = asyncio.Semaphore(CONCURRENCY)
+    return _sem
+
+
 async def synth(text: str, voice: str, out: str) -> None:
     import edge_tts
-    await edge_tts.Communicate(text, voice).save(out)
-    print(f"  {out}  ←  \"{text}\"")
+    from edge_tts.exceptions import NoAudioReceived
+
+    async with _semaphore():
+        for attempt in range(1, MAX_RETRIES + 1):
+            try:
+                await edge_tts.Communicate(text, voice).save(out)
+                print(f"  {out}  ←  \"{text}\"")
+                return
+            except NoAudioReceived:
+                if attempt == MAX_RETRIES:
+                    raise
+                await asyncio.sleep(RETRY_BACKOFF_SEC * attempt)
 
 
 async def main() -> None:

web/static/audio/manifest.json

@@ -13,4 +13,4 @@
       "gender": "male"
     }
   ]
-}
+}