Move Hume Octave TTS calls to Cartesia

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

Deadline (API shutdown): November 13, 2026. Hume's archived developer docs say the TTS and EVI APIs stay fully supported until this cutoff and end at 12:01 a.m. EST. Save the source recordings behind your cloned voices, and any generated audio you still need, while the account exists.

November 2, 2026: Hume's docs move to Audio Decisions. Hume's changelog announces the Audio Decisions API, and the old Octave documentation URLs now redirect to it. Keep your own copy of any TTS reference you still use. (https://dev.hume.ai/changelog)

## 1. Find the calls
Search for `api.hume.ai` or `HUME_API_KEY`. The request to replace is `POST /v0/tts` with an `utterances` array. 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: Hume Octave sends the key in `X-Hume-Api-Key` header from `HUME_API_KEY`. Replace it with `X-API-Key` from `CARTESIA_API_KEY` on the Cartesia path only, and keep `HUME_API_KEY` for the Hume Octave path. Hume keys stop working at the November 13 cutoff.
- Add `CARTESIA_API_KEY` wherever the project documents its environment (`.env.example`, README, deploy config). Never hardcode the key.
- Concurrency: Hume Octave allows Hume capped TTS requests per minute: 15 on Free and Starter, 75 on Creator and Pro, 150 on Scale and 225 on Business. 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 Hume Octave 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: Empathic support agent, female → `d7bf7d75-64b7-4c1e-86c0-79d647366587` (Michelle - Empathetic Voice); Empathic support agent, male → `7d444628-dd13-442b-b687-71a6baf0c07e` (Joseph - Empathetic Voice); Expressive narration and character reads → `083de431-6b5c-4b18-a2dc-264eafa205f2` (Diana - Animated Narrator); Calm wellness or meditation content → `3ccc4544-84f7-45e3-ae57-5c52b5a1fac6` (Aiden - Yogi).

## 3. Field mapping, Hume Octave → Cartesia
- `POST /v0/tts` with an `utterances` array → `POST /tts/bytes`, one request per utterance. For a continuing conversation, use one WebSocket context.
- `X-Hume-Api-Key` header → `X-API-Key` and `Cartesia-Version: 2026-08-14`
- `utterances[].text` → `transcript`
- `utterances[].voice` by `name` or `id` → `voice`: `{"id": "<voice id>"}`. Hume voices don't export, so you rebuild each one on Cartesia.
- `utterances[].description` (acting instructions or voice design) → `generation_config.emotion` plus a chosen voice. Free-text descriptions have no equivalent. Pick the closest voice and an emotion, then test on your scripts.
- `utterances[].speed` → `generation_config.speed` (0.6 to 1.5)
- `utterances[].trailing_silence` → A `<break>` tag at the end of the `transcript`
- `context`: earlier utterances or a generation to continue from → One WebSocket `context_id` with `continue: true`. `/tts/bytes` has no context field, so each request starts fresh.
- `include_timestamp_types`: `word`, `phoneme` → TTS WebSocket with `add_timestamps` or `add_phoneme_timestamps`. `/tts/bytes` returns audio with no timestamps.
- `format`: `mp3`, `wav` or `pcm` → `output_format` with `container` `mp3`, `wav` or `raw`
- `generations[].audio`, base64 in JSON → The response body is the audio. Drop the base64 decode.

## 4. Hume Octave fields with no Cartesia equivalent
- `num_generations`: 1 to 5 takes per request. What to do: Send the request again for another take.

## 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 = "I'm sorry that happened. Let's get it sorted out together."
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",
        "generation_config": {"emotion": "calm"},
        "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 Hume Octave or Cartesia per request from one config flag, defaulting to Hume Octave.
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 Hume Octave 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
- Hume developer docs (archived Oct 3, 2026): https://web.archive.org/web/20261003165712/https://dev.hume.ai/intro
- Hume TTS synthesize JSON reference (archived Apr 15, 2026): https://web.archive.org/web/20260415072845/https://dev.hume.ai/reference/text-to-speech-tts/synthesize-json
- Hume changelog: https://dev.hume.ai/changelog
- 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.