Move Rime AI's Coda API to Cartesia

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

August 15, 2026: Arcana retired from the Rime cloud API. Requests that set `modelId` to `arcana`, `arcanav2` or `arcanav3` are now served by Coda, which has a different voice catalog. Rime warns that an unsupported speaker and language pair can still return audio, so a successful response doesn't prove the pairing is supported. (https://docs.rime.ai/docs/arcana-sunset)

## 1. Find the calls
Search for `coda.api.rime.ai` or `RIME_API_KEY`. The request to replace is `POST https://coda.api.rime.ai/`. 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: Rime sends the key in `Authorization: Bearer <key>` from `RIME_API_KEY`. Replace it with `X-API-Key` from `CARTESIA_API_KEY` on the Cartesia path only, and keep `RIME_API_KEY` for the Rime path. Rime rejects a token sent without the `Bearer ` prefix as `invalid api key`. Cartesia's key goes in its own header with no prefix.
- Add `CARTESIA_API_KEY` wherever the project documents its environment (`.env.example`, README, deploy config). Never hardcode the key.
- Request size: Rime caps a request at 1,000 characters per request on the Coda HTTP endpoint. 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.
- Voice IDs don't carry over. Keep each voice ID in one named constant or config value, put a placeholder where each Rime 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: astra (young American woman, bright and quick) → `8d8ce8c9-44a4-46c4-b10f-9a927b99a853` (Connie - Candid Conversationalist); eyre (American woman, 30 to 50, calm and friendly) → `d7bf7d75-64b7-4c1e-86c0-79d647366587` (Michelle - Empathetic Voice); masonry (man, 30 to 50, confident and low) → `228fca29-3a0a-435c-8728-5cb483251068` (Kiefer - Assured Tone); godfrey (young American man, easygoing) → `630ed21c-2c5c-41cf-9d82-10a7fd668370` (Corey - Supportive Buddy).

## 3. Field mapping, Rime → Cartesia
- `POST https://coda.api.rime.ai/` → `POST https://api.cartesia.ai/tts/bytes`. Rime retired Arcana on August 15, 2026 and now serves Arcana calls with Coda, so Coda is the request to port.
- `Authorization: Bearer <key>` → `X-API-Key` and `Cartesia-Version: 2026-08-14`
- `text`, up to 1,000 characters → `transcript`
- `speaker`, such as `astra` → `voice`: `{"id": "<voice id>"}`
- `lang`, such as `en` → `language`
- `Accept: audio/wav`, `audio/mpeg`, `audio/PCMU`, `audio/L16` → `output_format`: `wav`, `mp3`, `raw` with `pcm_mulaw`, or `raw` with `pcm_s16le`. `audio/L16` is headerless 16-bit PCM, which Cartesia calls the `raw` container with `pcm_s16le`.
- `samplingRate` → `output_format.sample_rate`
- `timeScaleFactor` (below 1.0 is faster) → `generation_config.speed` (above 1.0 is faster). The two scales point in opposite directions. Cartesia's range is also narrower, 0.6 to 1.5 against Rime's 0.4 to 2.5.
- `customLexicon` (X-SAMPA, for `en`, `de`, `es`, `fr`, `it`, `pt`) → `pronunciation_dict_id`. Cartesia pronunciations are written in [IPA](https://docs.cartesia.ai/build-with-cartesia/capability-guides/phonemes), so convert each X-SAMPA entry before you load it into a dictionary.

## 4. Rime fields with no Cartesia equivalent
- `Accept`: Opus in WebM or Ogg. What to do: Ask for `mp3` or raw PCM and encode Opus yourself.

## 5. 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 = "Your table for four is booked for seven thirty on Friday."
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": 24000,
        },
    },
)
response.raise_for_status()
open("out.wav", "wb").write(response.content)
```

## 6. Roll out
1. Choose Rime or Cartesia per request from one config flag, defaulting to Rime.
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 Rime 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
- Rime Coda streaming HTTP reference: https://docs.rime.ai/api-reference/coda/http.md
- Rime Arcana sunset notice: https://docs.rime.ai/docs/arcana-sunset
- Rime voices: https://docs.rime.ai/docs/voices.md
- Rime API authentication: https://docs.rime.ai/docs/api-authentication.md
- Cartesia TTS bytes API reference: https://docs.cartesia.ai/api-reference/tts/bytes
- Cartesia IPA phonemes guide: https://docs.cartesia.ai/build-with-cartesia/capability-guides/phonemes

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