fix(rollout): handle invalid tool arguments

[MinhaoLi0318]

Summary

• Parse arguments for recognized SGLang and OpenAI tool-call envelopes before execution.
• Invoke the executor only when the decoded arguments are a JSON object.
• Return a deterministic tool-error response for malformed or non-object JSON so the rollout can continue and the model can correct the call on the next turn.
• Preserve existing behavior for valid objects, empty arguments, sampled assistant output/tokens, and envelope recognition.

Behavior

• JSON object: Call the executor with the parsed mapping and return the existing tool response.
• Empty string: Call the executor with an empty mapping and return the existing tool response.
• Malformed JSON: Do not call the executor; return a deterministic tool-error response.
• JSON list or number: Do not call the executor; return a deterministic tool-error response.

Tests

• Added parameterized regression coverage for both ToolCallItem and ChatCompletionMessageToolCall.
• Targeted regression cases: 10 passed.
• Pre-commit, Ruff, Black, isort, autoflake, and py_compile passed.
• The full CPU suite is left to PR CI because the local macOS environment does not contain the pinned SGLang/PyTorch source stack used by Miles CI.

Fixes #2689

@yueming-yuan

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[MinhaoLi0318]

@guapisolo @yueming-yuan please help take a look when you have time~