fix(openai): restore audio transcription multipart requests

[acedatacloud-dev]

Summary

• encode scalar multipart fields as an async-compatible mapping
• encode repeated array fields as multipart form parts without passing a synchronous tuple stream to httpx.AsyncClient
• replace the over-mocked unit test with a real httpx.MockTransport regression test

Production evidence

The Studio OpenAI MCP tool failed before sending the transcription request with:

Attempted to send an sync request with an AsyncClient instance.

The issue reproduced locally with the previous data=list[tuple] request shape. Both the original 34.9 MB recording and two valid sub-10 MB/sub-6 MB segments failed identically, ruling out file size, format, and URL access.

Verification

• python3 -m pytest -q (56 passed)
• python3 -m pytest tests/test_client.py tests/test_audio_tools.py -q (27 passed)
• python3 -m ruff check core/client.py tests/test_client.py
• python3 -m ruff format --check core/client.py tests/test_client.py
• python3 -m mypy core/client.py

Post-deploy

Re-run openai_transcribe_audio against one of the prepared audio segments and confirm the request reaches /v1/audio/transcriptions.

🤖 Generated with Claude Code