Fix render_lang_test: /v1/tts returns raw WAV, and add a model-portable voice mode

Two bugs that would have failed on all 120 clips:

- The script parsed {"audio_base64": ...}. /v1/tts returns raw WAV bytes
  with media_type audio/wav. Now validated with a RIFF check so a bad
  response says so instead of raising a JSON error.
- It preferred the cached ICL prompt. Those are MODEL-SPECIFIC: narrator.pt
  was built by the 1.7B model and the 0.6B rejects its ref_code with
  'Expected size 1024 but got size 2048'. --voice-mode now defaults to ref,
  sending ref_audio + ref_text (the Rainbow Passage the refs were generated
  from) so the model builds its own prompt; --voice-mode prompt is available
  and documented as valid only for the model that built the cache.

Measured on a reboot with swap at 0: the 1.7B model still fails to generate
(footprint 23.4G on a 16G machine, zero POSTs completing). The 0.6B model
loads in 5s and renders 6-21s per clip.

Rendered 120 clips across 10 languages plus a stitched file each: no silent
clips, RMS -22 to -23 dB throughout. Duration ordering is itself a useful
signal - ZH 92s and EN 81s against DE 98s, FR/IT 102-107s.

Suite: 721 passed, 8 pre-existing.

scripts/render_lang_test.py

@@ -10,10 +10,16 @@ per-language file, which is what you actually listen to.
 Qwen3-TTS auto-detects language from the text - ``/v1/tts`` takes no language
 argument - so there is nothing to set per clip.
 
-The speaker's cached ICL prompt is preferred over the raw reference WAV because
-it gives the identical voice on every utterance. Where no prompt is cached the
-script falls back to ``ref_audio`` + ``ref_text``, which needs the reference
-transcript to be correct or the clone is worse, not better.
+Voice is supplied two ways:
+
+``--voice-mode ref`` (default) sends ``ref_audio`` + ``ref_text`` and lets the
+model build its own prompt. ``--voice-mode prompt`` uses the cached ICL prompt,
+which is more faithful but is **model-specific**: a prompt built by the 1.7B
+model has a ref_code whose dimensions the 0.6B model rejects, with
+"Expected size 1024 but got size 2048". Cached prompts are therefore only safe
+when rendering with the model that built them.
+
+/v1/tts returns raw WAV bytes, not a JSON envelope.
 
 Usage:
     python scripts/render_lang_test.py --project DealHunterShip_4.0e_whole-to-audio
@@ -41,6 +47,10 @@ TTS_URL = "http://127.0.0.1:8881"
 CHAPTER_ID = 1
 TRANSLATION_DIRNAME = "Lang_test"
 
+# The reference WAVs were generated from this text, so it is the correct
+# ref_text for the clone path. A wrong transcript degrades the clone.
+from umbraaudio.voices.qwen_tts import VOICE_REFERENCE_TEXT as REF_TEXT  # noqa: E402
+
 
 def slugify(name: str) -> str:
     return name.strip().lower().replace(" ", "_")
@@ -124,14 +134,16 @@ def post_tts(
         headers={"Content-Type": f"multipart/form-data; boundary={boundary}"},
         method="POST",
     )
+    # Raw WAV, not JSON. An earlier version of this script parsed
+    # {"audio_base64": ...} and would have failed on every single clip.
     with urllib.request.urlopen(req, timeout=timeout) as resp:
-        payload = json.loads(resp.read())
-    audio = payload.get("audio_base64")
-    if not audio:
-        raise RuntimeError(f"no audio in response: {str(payload)[:200]}")
-    import base64
-
-    return base64.b64decode(audio)
+        body = resp.read()
+    if not body.startswith(b"RIFF"):
+        raise RuntimeError(
+            f"expected WAV, got {body[:80]!r} "
+            f"(HTTP {getattr(resp, 'status', '?')})"
+        )
+    return body
 
 
 def stitch(parts: list[Path], out: Path, gap_ms: int = 400) -> None:
@@ -162,6 +174,11 @@ def main() -> int:
     ap.add_argument("--languages", help="comma-separated suffixes, default all found")
     ap.add_argument("--timeout", type=int, default=900)
     ap.add_argument("--dry-run", action="store_true")
+    ap.add_argument(
+        "--voice-mode", choices=("ref", "prompt"), default="ref",
+        help="ref: ref_audio+ref_text (model-portable). prompt: cached ICL "
+             "prompt (more faithful, but only valid for the model that built it)",
+    )
     args = ap.parse_args()
 
     config = load_config(args.project)
@@ -231,11 +248,12 @@ def main() -> int:
                 continue
             t0 = time.time()
             try:
+                use_prompt = args.voice_mode == "prompt" and slug in prompts
                 audio = post_tts(
                     TTS_URL, u["text"],
-                    prompt_slug=slug if slug in prompts else None,
-                    ref_audio=None if slug in prompts else refs[slug],
-                    ref_text=None if slug in prompts else None,
+                    prompt_slug=slug if use_prompt else None,
+                    ref_audio=None if use_prompt else refs[slug],
+                    ref_text=None if use_prompt else REF_TEXT,
                     timeout=args.timeout,
                 )
                 dest.write_bytes(audio)