test(litellm): cover Responses named tool choice (#2953)

[ijevin]

Problem

Hindsight sends named tool choices to LiteLLM using the Chat Completions shape:

{"type":"function","function":{"name":"recall"}}

Responses-compatible endpoints instead require the function name at the top level (#2953).

Root cause

Hindsight correctly passes the canonical Chat Completions shape to litellm.acompletion(). Since #2950, the API requires LiteLLM 1.93.0 or newer; that version already includes the upstream Responses bridge normalization from BerriAI/litellm#27634, which rewrites the named choice before sending it to /responses.

Adding a second serializer in Hindsight would duplicate upstream routing logic and risk the two implementations drifting.

Change

Add focused regression coverage for both wire formats:

• A Responses-mode GitHub Copilot model reaches litellm.aresponses() with {"type":"function","name":"recall"}.
• A Chat Completions model retains {"type":"function","function":{"name":"recall"}}.

The Responses test exercises Hindsight's real call_with_tools() -> litellm.acompletion() -> Responses bridge path, mocking only the downstream provider request.

Validation

uv run pytest -o addopts='' tests/test_litellm_tool_choice.py tests/test_codex_tool_choice.py -q
4 passed

uv run ruff check .
All checks passed!

uv run ruff format --check .
568 files already formatted

Closes #2953.