Migrate Inworld TTS to Cartesia

Move this codebase from Inworld TTS to Cartesia for text to speech. Keep Inworld TTS working behind a flag until I switch over.

June 15, 2026: `inworld-tts-1` and `inworld-tts-1-max` discontinued. Inworld now sends requests for both ids to newer models, so audio you tuned on TTS 1 or TTS 1 Max now comes from a model you never tested. Inworld also marks `inworld-tts-1.5-max` and `inworld-tts-1.5-mini` as deprecated, with no end date. (https://docs.inworld.ai/tts/tts-models)

## 1. Find the calls
Search for `api.inworld.ai` or `INWORLD_API_KEY`. The request to replace is `POST https://api.inworld.ai/tts/v1/voice`. List every call site before you edit anything.

## 2. Target
- Request: `POST https://api.cartesia.ai/tts/bytes`
- Model: `model_id: "sonic-3.6"`
- Headers: `X-API-Key` read from the `CARTESIA_API_KEY` environment variable, and `Cartesia-Version: 2026-08-14`
- Auth swap: Inworld TTS sends the key in `Authorization: Basic <base64 key_id:secret>` from `INWORLD_API_KEY`. Replace it with `X-API-Key` from `CARTESIA_API_KEY` on the Cartesia path only, and keep `INWORLD_API_KEY` for the Inworld TTS path. Inworld's portal copies the credential already Base64-encoded, so it goes in as is. Cartesia keys need no encoding.
- Add `CARTESIA_API_KEY` wherever the project documents its environment (`.env.example`, README, deploy config). Never hardcode the key.
- Request size: Inworld TTS caps a request at 2,000 characters on the synthesize endpoint; 4,000 on HTTP streaming. Cartesia publishes no per-request maximum, so keep any chunking the code already does and split at sentence boundaries; for long text, stream over the Cartesia TTS WebSocket.
- Concurrency: Inworld TTS allows 5 concurrent requests on On-Demand, 10 on Creator, 50 on Builder, 150 on Developer, 500 on Growth. Look up the Cartesia plan's concurrent request limit at https://www.cartesia.ai/pricing, cap parallel requests at it, and retry rate-limit errors with backoff.
- Voice IDs don't carry over. Keep each voice ID in one named constant or config value, put a placeholder where each Inworld TTS voice was, and list them for me so I can pick replacements at https://www.cartesia.ai/voices.
- Suggested starting points to put in those placeholders, for me to confirm by ear: A warm voice for a companion or social app → `ec1e269e-9ca0-402f-8a18-58e0e022355a` (Ariana - Kind Friend); A deep voice for a guard, villain or other game character → `6a176356-ada1-4b48-b2ae-3a3fdd485680` (Elias - Night Warden); A narrator or oracle with gravitas for quests and cutscenes → `d7862948-75c3-4c7c-ae28-2959fe166f49` (Caspian - Oracle); A tutor or language partner → `f9fc912e-52f0-448a-8bfa-47e9ca75f25a` (Marilyn - Explainer).

## 3. Field mapping, Inworld TTS → Cartesia
- `POST https://api.inworld.ai/tts/v1/voice` → `POST https://api.cartesia.ai/tts/bytes`
- `Authorization: Basic <base64 key_id:secret>` → `X-API-Key` and `Cartesia-Version: 2026-08-14`
- `text`, up to 2,000 characters (4,000 on HTTP streaming) → `transcript`
- `modelId`: `inworld-tts-2`, `inworld-tts-2-flash`, or a deprecated `inworld-tts-1.5-max` or `inworld-tts-1.5-mini` → `model_id`: `sonic-3.6`. Inworld now routes `inworld-tts-1` and `inworld-tts-1-max` to newer models, so code that still names them is already running on a different model.
- `voiceId`, such as `Ashley` → `voice`: `{"id": "<voice id>"}`
- `language` (BCP-47) → `language`: `en` or `en-GB`
- `audioConfig.audioEncoding`: `LINEAR16` → `output_format`: `wav` container, `pcm_s16le`. Inworld's `LINEAR16` carries a WAV header, so `wav` is the match. Cartesia outputs `wav`, `mp3` and `raw` PCM only, so transcode for Opus or anything else.
- `audioConfig.sampleRateHertz` → `output_format.sample_rate`
- `audioConfig.speakingRate` (0.5 to 1.5) → `generation_config.speed` (0.6 to 1.5). Cartesia's floor is 0.6, so send 0.6 for any Inworld rate below it.
- `timestampType`: `WORD` or `CHARACTER` → `add_timestamps` on the TTS WebSocket. Word timestamps arrive as WebSocket events, and `/tts/bytes` returns none. Character timings and visemes have no equivalent.
- `pronunciationDictionarySettings` → `pronunciation_dict_id`. One dictionary per request, on `sonic-3` and newer.
- Response `audioContent` (base64) → Response body is the audio bytes

## 4. Reference request
Match the language and HTTP client the codebase already uses. This is the same request in Python:
```python
import os
import requests

text = "Stay close, traveler. The bridge ahead won't hold us both."
VOICE_ID = "6ccbfb76-1fc6-48f7-b71d-91ac6298247b"  # from /voices

response = requests.post(
    "https://api.cartesia.ai/tts/bytes",
    headers={
        "X-API-Key": os.environ["CARTESIA_API_KEY"],
        "Cartesia-Version": "2026-08-14",
    },
    json={
        "model_id": "sonic-3.6",
        "transcript": text,
        "voice": {"id": VOICE_ID},
        "language": "en",
        "output_format": {
            "container": "wav",
            "encoding": "pcm_s16le",
            "sample_rate": 44100,
        },
    },
)
response.raise_for_status()
open("out.wav", "wb").write(response.content)
```

## 5. Roll out
1. Choose Inworld TTS or Cartesia per request from one config flag, defaulting to Inworld TTS.
2. Run the same scripts through both and save the clips side by side for me to listen to. Check that the container, encoding and sample rate match what downstream code expects.
3. Leave the Inworld TTS client, call and key in place so rollback is flipping the flag. Don't delete them in this change.
4. Run the existing tests, and add one that checks the Cartesia request body.

## Docs
- Inworld synthesize speech reference: https://docs.inworld.ai/api-reference/ttsAPI/texttospeech/synthesize-speech
- Inworld TTS models: https://docs.inworld.ai/tts/tts-models
- Inworld API key authentication: https://docs.inworld.ai/portal/api-key-auth
- Cartesia TTS bytes API reference: https://docs.cartesia.ai/api-reference/tts/bytes

When you're done, tell me which files changed, which fields you couldn't map, and how to flip the flag.