Move Deepgram Aura-2 TTS to Cartesia

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

## 1. Find the calls
Search for `api.deepgram.com` or `DEEPGRAM_API_KEY`. The request to replace is `POST https://api.deepgram.com/v1/speak`. 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: Deepgram sends the key in `Authorization: Token <key>` from `DEEPGRAM_API_KEY`. Replace it with `X-API-Key` from `CARTESIA_API_KEY` on the Cartesia path only, and keep `DEEPGRAM_API_KEY` for the Deepgram path. Deepgram also accepts a short-lived token sent as `Bearer`.
- Add `CARTESIA_API_KEY` wherever the project documents its environment (`.env.example`, README, deploy config). Never hardcode the key.
- Request size: Deepgram caps a request at 2,000 characters per request on Aura-2 and Aura-1. 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: Deepgram allows 15 concurrent REST requests on Pay as you go and Growth; 45 and 60 streaming. 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 Deepgram 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: aura-2-thalia-en (clear, energetic; customer service and IVR) → `f786b574-daa5-4673-aa0c-cbe3e8534c02` (Katie - Friendly Fixer); aura-2-andromeda-en (casual, expressive; customer service) → `9626c31c-bec5-4cca-baa8-f8ba9e84c8bc` (Jacqueline - Reassuring Agent); aura-2-apollo-en (confident, casual) → `86e30c1d-714b-4074-a1f2-1cb6b552fb49` (Carson - Curious Conversationalist); aura-2-orion-en (calm, polite; informative) → `47c38ca4-5f35-497b-b1a3-415245fb35e1` (Daniel - Modern Assistant).

## 3. Field mapping, Deepgram → Cartesia
- `POST https://api.deepgram.com/v1/speak` → `POST /tts/bytes`
- `Authorization: Token <key>` → `X-API-Key` and `Cartesia-Version: 2026-08-14`
- `model=aura-2-thalia-en` (model, voice and language in one name) → `model_id`: `sonic-3.6`, `voice`: `{"id": "<voice id>"}` and `language`: `en`. Split the name in three. The voice becomes an ID and the suffix becomes `language`.
- `text` in the JSON body (2,000 characters at most) → `transcript`
- `encoding`: `linear16`, `mulaw` or `alaw` → `output_format.encoding`: `pcm_s16le`, `pcm_mulaw` or `pcm_alaw`
- `container`: `wav` or `none` → `output_format.container`: `wav` or `raw`
- `encoding`: `mp3`, with `bit_rate` → `container`: `mp3`, with `bit_rate` and `sample_rate`. Deepgram fixes MP3 at 22050 Hz. Cartesia takes 8000 to 48000.
- `sample_rate` (default 24000) → `output_format.sample_rate`
- `speed` (0.7 to 1.5) → `generation_config.speed` (0.6 to 1.5)
- Inline IPA pronunciation objects in `text` → `pronunciation_dict_id` (one dictionary). Move each override into a Cartesia pronunciation dictionary and drop the inline objects from the text.
- `wss://api.deepgram.com/v1/speak` → TTS WebSocket. Set `add_timestamps` there to get word timings with each audio chunk.

## 4. Deepgram fields with no Cartesia equivalent
- `encoding`: `opus`, `flac` or `aac`. What to do: Request `mp3` or raw PCM and transcode.
- `callback` and `callback_method`. What to do: Call `/tts/bytes` from a background worker and post the audio to your own webhook.
- `tag`, `mip_opt_out`. What to do: Drop them. Keep your own request log if you used `tag` for reporting.

## 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 = "You're booked for Tuesday at 3 p.m. Press 1 to confirm or 2 to reschedule."
VOICE_ID = "f786b574-daa5-4673-aa0c-cbe3e8534c02"  # 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": "raw",
            "encoding": "pcm_mulaw",
            "sample_rate": 8000,
        },
    },
)
response.raise_for_status()
open("out.ulaw", "wb").write(response.content)
```

## 6. Roll out
1. Choose Deepgram or Cartesia per request from one config flag, defaulting to Deepgram.
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 Deepgram 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
- Deepgram speak request reference: https://developers.deepgram.com/reference/text-to-speech/speak-request
- Deepgram text-to-speech docs: https://developers.deepgram.com/docs/text-to-speech
- Deepgram TTS voice controls: https://developers.deepgram.com/docs/tts-voice-controls
- Deepgram TTS models: https://developers.deepgram.com/docs/tts-models
- 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.