[codex] add lazy multilingual celery runtime

MULTILINGUAL_BENCHMARK_NOTES.md

@@ -0,0 +1,43 @@
+# Chatterbox Multilingual Runtime And Benchmark Notes
+
+Updated: 2026-05-15
+
+## Runtime shape
+
+- API: `http://127.0.0.1:8010`
+- Endpoint: `POST /v1/audio/speech`
+- Queue: `chatterbox_mtl_3060x2`
+- Workers: 2 total, one on physical GPU `2`, one on physical GPU `3`
+- GPU policy: RTX 3060 only, with `EXPECTED_GPU_NAME=RTX 3060`
+- Lazy behavior: API stays CPU-side; workers load on first request and unload after `300` seconds idle
+- VRAM gate: `MIN_FREE_VRAM_MB=4500` before model load
+- Chunking: Spanish sentence-boundary split, dynamic Celery dispatch, ordered PCM stitching in the API process
+- Result collection: chunk task results are polled with `ready()` and read from `task.result`, avoiding Celery Redis `task.get()` hangs on large fan-out
+
+## 100-sentence Spanish benchmark
+
+Test input: 100 Spanish dot-terminated sentences.
+
+| Metric | Value |
+| --- | ---: |
+| Audio duration | `541.180s` |
+| Client wall time | `173.295s` |
+| Client speed | `3.1229x` realtime |
+| Server speed | `3.1237x` realtime |
+| RTF | `0.3202` |
+| Chunks / tasks | `100 / 100` |
+| Worker split | `pid:1305231=52`, `pid:1305277=48` |
+| ASR speed on Parakeet `:5092/v1` | `22.5024x` realtime |
+| ASR WER | `1.0200` |
+
+Output files:
+
+- `/home/op/tts_unified_benchmark_outputs/chatterbox_multilingual_100_sentences.wav`
+- `/home/op/tts_unified_benchmark_outputs/chatterbox_multilingual_100_sentences_asr.txt`
+- `/home/op/tts_unified_benchmark_outputs/unified_tts_100_sentence_benchmark_results.json`
+
+## Notes
+
+- Distribution across the two workers is balanced; model generation is the bottleneck.
+- This path is slower than Turbo on the same 3060 pair but now has the same lazy Celery behavior and VRAM gate.
+- vLLM2 correctly refused to load immediately after this benchmark while this model was resident and free VRAM was below its configured gate.