fix(agent): preserve malformed tool-call arguments

[primorLee]

Summary

• preserve the model-produced tool argument string alongside its parsed mapping
• return the original string unchanged as OpenAI function.arguments
• keep the parsed mapping for chat-template replay and trajectory matching
• fail closed in the Anthropic adapter when arguments cannot be represented as an input object
• mark valid non-object JSON arguments as ill-formed instead of emitting invalid tool inputs

Problem

When SGLang recognizes a tool call but its arguments are malformed JSON, the parser currently replaces the argument text with a {"_raw_arguments": ...} sentinel mapping.

The OpenAI adapter then serializes that sentinel, so clients receive a different argument string from the one produced by the model. The Anthropic adapter emits the sentinel as a normal executable tool_use.input, even though Anthropic requires an input object and the model output was not valid.

For the model output:

{"cmd": "pytest"

current main returns:

OpenAI: {"_raw_arguments": "{\"cmd\": \"pytest\""}
Anthropic: tool_use(input={"_raw_arguments": ...})

This PR keeps two representations at the parsing boundary: the original wire text and the canonical mapping used for history rendering. OpenAI receives the former unchanged. Anthropic only exposes calls whose arguments are representable as an object; an invalid call becomes a normal non-tool response and the turn remains marked ill_formed.

Valid object-valued tool calls and XML fallback calls keep their existing behavior.

Validation

• full CPU agent suite: 70 passed, 1 skipped
• changed-file pre-commit: case-conflict, private-key, large-file, Ruff, autoflake, isort, and Black hooks passed
• Python compilation and git diff --check passed
• focused coverage includes malformed JSON, valid raw formatting, non-object JSON, OpenAI wire preservation, and Anthropic fail-closed behavior

Closes #2299