Coqui TTS alternative: move XTTS-v2 to an API

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

Company shut down on January 2024. Coqui, the company behind the `TTS` library and XTTS-v2, announced in early January 2024 that it was shutting down. The original `coqui-ai/TTS` repository last shipped a release, v0.22.0, on December 12, 2023. Idiap Research Institute maintains a fork published on PyPI as `coqui-tts`.

December 12, 2023: Last release of coqui-ai/TTS. v0.22.0 is the final release in Coqui's own repository. Fixes since then ship only in the Idiap fork, so a pinned `TTS` install gets no patches. (https://github.com/coqui-ai/TTS/releases)

## 1. Find the calls
Search for `TTS.api`. The request to replace is `TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda")`. 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: Coqui TTS sends the key in None; the library runs inside your own process. Replace it with `X-API-Key` from `CARTESIA_API_KEY` on the Cartesia path only, and keep Coqui TTS's key for the Coqui TTS path. Moving to an API adds a secret to manage. Keep it server-side.
- Add `CARTESIA_API_KEY` wherever the project documents its environment (`.env.example`, README, deploy config). Never hardcode the key.
- Voice IDs don't carry over. Keep each voice ID in one named constant or config value, put a placeholder where each Coqui 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: Friendly assistant or app prompts → `db6b0ed5-d5d3-463d-ae85-518a07d3c2b4` (Skylar - Friendly Guide); Calm male narration → `98a34ef2-2140-4c28-9c71-663dc4dd7022` (Clyde - Calm Narrator); Expressive storytelling → `a33f7a4c-100f-41cf-a1fd-5822e8fc253f` (Lauren - Lively Narrator).

## 3. Field mapping, Coqui TTS → Cartesia
- `TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda")` → `POST /tts/bytes`. Nothing loads in your process. Each synthesis is one HTTPS request.
- No auth (the model runs in your own process) → `X-API-Key` and `Cartesia-Version: 2026-08-14`. Keep the key on your server, in `CARTESIA_API_KEY`.
- Model name `xtts_v2` → `model_id`: `sonic-3.6`
- `text` → `transcript`
- `speaker_wav`: a reference clip sent with every call → `voice`: `{"id": "<voice id>"}`. Clone the same clip once with `client.voices.clone(clip=..., language=..., name=...)` and store the ID it returns.
- `speaker`: a built-in XTTS speaker name → `voice`: `{"id": "<voice id>"}` from the voice library. Built-in XTTS speakers do not exist on Cartesia. Pick a library voice by role.
- `language`, one of 17 codes such as `en` or `zh-cn` → `language`, such as `en` or `en-GB`. Check each code you send against Cartesia's [language list](/languages).
- `file_path="out.wav"` → `output_format` with `container: wav`. The API returns audio bytes. Your code writes the file.
- `speed` (XTTS inference, default 1.0) → `generation_config.speed` (0.6 to 1.5)
- `model.inference_stream(...)` for chunked audio → TTS WebSocket. `/tts/bytes` also streams its response body as audio is generated.
- A fine-tuned XTTS checkpoint → Professional voice clone. Checkpoints do not transfer. A professional clone trains on 30 minutes of audio and needs a Startup plan or higher.

## 4. Coqui TTS fields with no Cartesia equivalent
- `temperature`, `top_k`, `top_p`, `repetition_penalty`, `length_penalty`. What to do: Drop them. Sonic 3.6 exposes no sampling parameters; choose the voice and emotion by ear.

## 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 = "Welcome back. Your next lesson starts in two minutes."
VOICE_ID = "6ccbfb76-1fc6-48f7-b71d-91ac6298247b"  # from /voices, or your clone

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 Coqui TTS or Cartesia per request from one config flag, defaulting to Coqui 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 Coqui 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
- coqui-ai/TTS README: https://github.com/coqui-ai/TTS
- Idiap coqui-ai-TTS fork: https://github.com/idiap/coqui-ai-TTS
- XTTS documentation (Idiap fork): https://coqui-tts.readthedocs.io/en/latest/models/xtts.html
- XTTS-v2 model card: https://huggingface.co/coqui/XTTS-v2
- Cartesia TTS bytes API reference: https://docs.cartesia.ai/api-reference/tts/bytes
- Cartesia voice cloning guide: https://docs.cartesia.ai/build-with-cartesia/capability-guides/clone-voices

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