OpenAI Whisper alternative: migrate to Cartesia

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

Deadline (Model removal): February 26, 2027. OpenAI removes `whisper-1`, `gpt-4o-transcribe`, `gpt-4o-mini-transcribe` and `gpt-4o-transcribe-diarize` from its API on this date. Calls that name any of the four start failing.

August 26, 2026: OpenAI deprecates whisper-1 and its gpt-4o transcription models. OpenAI notified developers that the four transcription models would leave the API in six months. It recommends `gpt-transcribe` for files and `gpt-live-transcribe` for live audio. (https://developers.openai.com/api/docs/deprecations)

## 1. Find the calls
Search for `openai` or `OPENAI_API_KEY`. The request to replace is `POST /v1/audio/transcriptions`. List every call site before you edit anything.

## 2. Target
- Recorded files: `client.stt.transcribe(...)` in the `cartesia` SDK, or `POST https://api.cartesia.ai/stt`, with `model: "ink-whisper"`
- Live audio: `ink-2` over the STT WebSocket
- Headers: `X-API-Key` read from the `CARTESIA_API_KEY` environment variable, and `Cartesia-Version: 2026-08-14`
- Auth swap: OpenAI Whisper API 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 Whisper API path. The open-source Whisper package runs on your own machine and needs no API key.
- Add `CARTESIA_API_KEY` wherever the project documents its environment (`.env.example`, README, deploy config). Never hardcode the key.
- Request size: OpenAI Whisper API caps a request at 25 MB per uploaded file, as mp3, mp4, mpeg, mpga, m4a, wav or webm. Cartesia publishes no per-request maximum, so keep any file splitting the code already does until larger files are tested.
- Always pass `language`. Cartesia's batch call doesn't detect it and assumes `en`.

## 3. Field mapping, OpenAI Whisper API → Cartesia
- `POST /v1/audio/transcriptions` → `client.stt.transcribe(...)` in the Cartesia SDK, or `POST /stt`
- `Authorization: Bearer $OPENAI_API_KEY` → `Cartesia(api_key=...)`
- `file` (multipart upload, up to 25 MB) → `file`, an open binary file. Batch STT accepts uploads only. Download remote audio before you send it.
- `model`: `whisper-1` → `model`: `ink-whisper`
- `language`, optional → `language`. Whisper detects the language when you leave it out. Cartesia's batch call assumes `en` and does not detect, so set it every time.
- `timestamp_granularities[]`: `word` → `timestamp_granularities`: `["word"]`. The response carries `words`, each with `word`, `start` and `end`. Segment-level timestamps have no equivalent.
- Open-source `whisper`: `model.transcribe(..., word_timestamps=True)` → `timestamp_granularities`: `["word"]`. Self-hosted `word_timestamps=True` and the API's word granularity both become this one argument on Cartesia.
- Realtime transcription session → `client.stt.auto_finalize.websocket(model="ink-2", ...)`. Ink-2 handles English, Spanish, French, Hindi and Japanese, and detects which one it hears.

## 4. OpenAI Whisper API fields with no Cartesia equivalent
- `prompt`: vocabulary or style hint. What to do: Ink-2 on the WebSocket takes up to 100 `keyterm` values. The batch endpoint takes none.
- `response_format`: `srt` or `vtt`. What to do: Build caption cues from the `words` timestamps.
- `gpt-4o-transcribe-diarize` speaker labels. What to do: Cartesia STT returns no speaker labels. Keep a separate diarization step if you depend on them.
- `POST /v1/audio/translations`. What to do: Transcribe in the source language, then translate the text with your LLM.

## 5. Reference request
Match the language and HTTP client the codebase already uses. This is the same request in Python:
```python
import os

from cartesia import Cartesia

client = Cartesia(api_key=os.environ["CARTESIA_API_KEY"])

with open("meeting.mp3", "rb") as audio_file:
    result = client.stt.transcribe(
        file=audio_file,
        model="ink-whisper",
        language="en",
        timestamp_granularities=["word"],
    )

print(result.text)
for word in result.words or []:
    print(f"{word.start:6.2f}s  {word.word}")
```

## 6. Roll out
1. Choose OpenAI Whisper API or Cartesia per request from one config flag, defaulting to OpenAI Whisper API.
2. Transcribe the same recordings with both and save the transcripts side by side. Compare the words, the word timestamps and the language on each.
3. Leave the OpenAI Whisper 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
- OpenAI deprecations: https://developers.openai.com/api/docs/deprecations
- OpenAI speech-to-text guide: https://developers.openai.com/api/docs/guides/speech-to-text
- OpenAI API pricing: https://developers.openai.com/api/docs/pricing
- Open-source Whisper on GitHub (MIT License, word_timestamps): https://github.com/openai/whisper
- Cartesia speech-to-text docs: https://docs.cartesia.ai/use-the-api/stt/compare-endpoints

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