[codex] Add truncated reasoning response error

[rasdani]

Summary

Adds a narrow TruncatedReasoningError for renderer responses that hit finish_reason == "length" after emitting reasoning but before producing content or tool calls.

The new error subclasses EmptyModelResponseError, so existing retry/catch behavior remains compatible while downstream logs can distinguish true empty responses from length-truncated reasoning-only responses.

Validation

• uv run ruff format verifiers/errors.py verifiers/clients/renderer_client.py verifiers/utils/error_utils.py tests/test_renderer_client.py tests/test_error_chain.py
• uv run ruff check verifiers/errors.py verifiers/clients/renderer_client.py verifiers/utils/error_utils.py tests/test_renderer_client.py tests/test_error_chain.py
• uv run pytest tests/test_renderer_client.py::test_renderer_client_rejects_reasoning_only_native_response tests/test_renderer_client.py::test_renderer_client_rejects_truncated_reasoning_native_response tests/test_error_chain.py::TestErrorChain::test_truncated_reasoning_error_round_trips
• Replayed saved Nemotron Nano SWE repro with PYTHONPATH=/root/git/verifiers; the capped response now classifies as TruncatedReasoningError with finish_reason=length, content_len=0, tool_call_count=0, and reasoning_len=611.

Note

Add TruncatedReasoningError for responses that hit length limit during reasoning

• Adds TruncatedReasoningError as a subclass of EmptyModelResponseError in verifiers/errors.py to distinguish length-truncated reasoning responses from generic empty responses.
• Updates RendererClient.raise_from_native_response in renderer_client.py to raise TruncatedReasoningError instead of EmptyModelResponseError when a response has reasoning_content, no content or tool calls, and finish_reason == 'length'.
• Registers TruncatedReasoningError in vf_error_types so it round-trips correctly through error serialization.

^{Macroscope summarized 6808f92.}