Migrate the OpenAI TTS API to Cartesia

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

Deadline (Shutdown announced): January 6, 2027. OpenAI's text-to-speech guide gives January 6, 2027 as the shutdown date for `tts-1`, `tts-1-hd` and `gpt-4o-mini-tts`, including its `2025-03-20` and `2025-12-15` snapshots. Existing Speech API integrations keep working until that day.

October 1, 2026: OpenAI deprecates its text-to-speech models. OpenAI's deprecation notice marks `tts-1`, `tts-1-hd` and every `gpt-4o-mini-tts` snapshot as deprecated, with at least three months' notice before removal. (https://developers.openai.com/api/docs/deprecations#2026-10-01-text-to-speech-models)

## 1. Find the calls
Search for `api.openai.com` or `OPENAI_API_KEY`. The request to replace is `POST /v1/audio/speech`. 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: OpenAI TTS sends the key in `Authorization: Bearer` header from `OPENAI_API_KEY`. Replace it with `X-API-Key` from `CARTESIA_API_KEY` on the Cartesia path only, and keep `OPENAI_API_KEY` for the OpenAI TTS path.
- Add `CARTESIA_API_KEY` wherever the project documents its environment (`.env.example`, README, deploy config). Never hardcode the key.
- Request size: OpenAI TTS caps a request at 4,096 characters of `input` per request. 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 OpenAI 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: `marin` or `cedar` on a support assistant → `f786b574-daa5-4673-aa0c-cbe3e8534c02` (Katie - Friendly Fixer); Order, delivery and account notifications → `25d7abcb-4d6d-4aca-adce-8a1c85620c8b` (Jessica - Clear Communicator); An in-app digital assistant → `47c38ca4-5f35-497b-b1a3-415245fb35e1` (Daniel - Modern Assistant); Read-aloud for articles and long posts → `98a34ef2-2140-4c28-9c71-663dc4dd7022` (Clyde - Calm Narrator).

## 3. Field mapping, OpenAI TTS → Cartesia
- `POST /v1/audio/speech` → `POST /tts/bytes`. Use the WebSocket when you need audio before the text is complete.
- `Authorization: Bearer $OPENAI_API_KEY` → `X-API-Key` and `Cartesia-Version: 2026-08-14`. Requests without the version header fail.
- `model`: `gpt-4o-mini-tts`, `tts-1`, `tts-1-hd` → `model_id`: `sonic-3.6`
- `input` (up to 4,096 characters) → `transcript`
- `voice`: a built-in name such as `marin`, or a custom voice `{"id": "voice_..."}` → `voice`: `{"id": "<voice id>"}`. OpenAI's built-in names and custom `voice_...` IDs both stay on OpenAI. Pick a library voice, or clone again from the consented source recording.
- `instructions`: free-text style guidance → `generation_config`: `emotion`, `speed`, `volume`. Structured fields take the place of the prompt. Rewrite each instruction as an emotion and a speed, then listen.
- `speed`: 0.25 to 4.0 → `generation_config.speed`: 0.6 to 1.5. Clamp anything outside 0.6 to 1.5, since Cartesia has no setting for it.
- `response_format`: `mp3`, `wav`, `pcm` → `output_format`: a `container` of `mp3`, `wav` or `raw`, plus `encoding` and `sample_rate`. OpenAI's `pcm` is 24 kHz `pcm_s16le`. Set `raw`, `pcm_s16le` and `24000` to match it byte for byte.
- `stream_format`: `sse` or `audio` → TTS WebSocket. `/tts/bytes` streams the response body as it generates. The WebSocket also accepts text in pieces.

## 4. OpenAI TTS fields with no Cartesia equivalent
- `response_format`: `opus`, `aac`, `flac`. What to do: Request `mp3` or raw PCM and transcode to the container you need.

## 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 order has shipped and should arrive on Thursday."
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)
```

## 6. Roll out
1. Choose OpenAI TTS or Cartesia per request from one config flag, defaulting to OpenAI 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 OpenAI 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
- OpenAI text-to-speech guide: https://developers.openai.com/api/docs/guides/text-to-speech
- OpenAI deprecations: https://developers.openai.com/api/docs/deprecations
- OpenAI create speech API reference: https://developers.openai.com/api/reference/resources/audio/subresources/speech/methods/create
- Cartesia TTS bytes API reference: https://docs.cartesia.ai/api-reference/tts/bytes
- Cartesia TTS WebSocket API reference: https://docs.cartesia.ai/api-reference/tts/websocket

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