Move the Murf AI API to Cartesia

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

October 5, 2026: Gen2 streaming decommissioned. Murf took Gen2 off its streaming API and points streaming users to Falcon 2.1. Gen2 now runs only on the non-streaming generate endpoint, so any Gen2 streaming code needs rewriting whichever vendor you pick. (https://murf.ai/api/docs/changelog)

## 1. Find the calls
Search for `api.murf.ai` or `MURF_API_KEY`. The request to replace is `POST /v1/speech/generate` (Gen2). 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: Murf API sends the key in `api-key` header from `MURF_API_KEY`. Replace it with `X-API-Key` from `CARTESIA_API_KEY` on the Cartesia path only, and keep `MURF_API_KEY` for the Murf API path. Murf also accepts a short-lived `token` header minted from its generate-token endpoint. Cartesia takes the API key on server-side calls.
- Add `CARTESIA_API_KEY` wherever the project documents its environment (`.env.example`, README, deploy config). Never hardcode the key.
- Request size: Murf API caps a request at 3,000 characters 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.
- Concurrency: Murf API allows 5 concurrent TTS requests on the free trial and pay as you go; Falcon 2.1 allows 2 on the global and regional endpoints, 5 on US-East. 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 Murf API 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: en-US-natalie (US English, e-learning and explainers) → `26403c37-80c1-4a1a-8692-540551ca2ae5` (Marian - Poised Narrator); en-US-miles (US English, instructional) → `2948c301-9211-4112-8f36-4c3fc836ef12` (Bryce - Clear Explainer); en-UK-ruby (British English, calm voiceover) → `273f9ef7-9fc2-4def-88bb-ab108c6249ca` (Julia - Gentle Guide); en-UK-theo (British English, narration) → `5568a7df-e5ab-4442-9fae-2e9ba1b15ad8` (Quentin - Refined Narrator).

## 3. Field mapping, Murf API → Cartesia
- `POST /v1/speech/generate` (Gen2) → `POST /tts/bytes`
- `POST /v1/speech/stream` or the `stream-input` WebSocket (Falcon 2.1) → The TTS WebSocket. Sonic sends audio back in chunks while it generates; the [WebSocket reference](https://docs.cartesia.ai/api-reference/tts/websocket) has the message format.
- `api-key` header → `X-API-Key` and `Cartesia-Version: 2026-08-14`
- `text`, up to 3,000 characters → `transcript`
- `voiceId`, such as `en-US-natalie` → `voice`: `{"id": "<voice id>"}`
- `modelVersion`: `GEN2` → `model_id`: `sonic-3.6`
- `locale`, such as `en-US` → `language`: `en` or `en-US`
- `rate` (-50 to 50) → `generation_config.speed`. Cartesia's speed is a multiplier from 0.6 to 1.5. The two scales don't convert exactly, so set it by ear.
- `format`: `WAV`, `MP3`, `PCM`, `ULAW`, `ALAW` → `output_format.container` and `encoding`. `ULAW` becomes `raw` with `pcm_mulaw`, `ALAW` becomes `raw` with `pcm_alaw`, and `PCM` becomes `raw` with `pcm_s16le`.
- `sampleRate` → `output_format.sample_rate`
- `pronunciationDictionary`, a map sent with each request → `pronunciation_dict_id`. Create the dictionary once on Cartesia, then pass its ID.
- `wordDurations` in the response → `add_timestamps` on the TTS WebSocket. `/tts/bytes` returns no timings. Word timestamps come back as WebSocket events.
- `audioFile` URL, or `encodedAudio` with `encodeAsBase64` → Audio bytes in the response body

## 4. Murf API fields with no Cartesia equivalent
- `pitch` (-50 to 50). What to do: Cast a voice whose natural pitch already fits the script.
- `format`: `FLAC`, `OGG`. What to do: Generate `wav` or `mp3` and convert it in your pipeline.

## 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 = "Module three covers how to file an expense report."
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 Murf API or Cartesia per request from one config flag, defaulting to Murf API.
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 Murf API 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
- Murf generate speech reference: https://murf.ai/api/docs/api-reference/text-to-speech/generate
- Murf stream speech reference: https://murf.ai/api/docs/api-reference/text-to-speech/stream
- Murf rate limits and concurrency: https://murf.ai/api/docs/resources/rate-limits
- 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.