SessionManager._parse_message repr-mangles non-str tool_result content, and ToolAnalyzer's error detection silently depends on that mangling

[Rahul-s-007]

Split out of PR #43 per review — this is a distinct benchmark-harness data-integrity bug, not the Unicode-detection feature itself.

Bug 1: non-string message content is corrupted before it's ever written to disk.

SessionManager._parse_message (Detection/main_benchmark.py) passes tool_result message content — which Claude Code's JSONL always carries as a list of content blocks — straight into _truncate_content:

def _truncate_content(self, content: str, max_length: int = 10000) -> str:
    if isinstance(content, str):
        return content[:max_length] + "..." if len(content) > max_length else content
    return str(content)[:max_length] + "..." if len(str(content)) > max_length else str(content)

The non-str branch calls str() on the list, which invokes Python's repr() on every element — silently escaping any non-printable Unicode into literal backslash text (e.g. a real hidden-Unicode payload becomes the 10 literal ASCII characters \U000e0069, not the actual codepoint) before json.dump ever writes claude_conversation.json. Irreversible: nothing downstream un-reprs it, including main_detector.py's _convert_conversation_to_messages, which reads that file back unmodified. In a real recorded benchmark run, this hits 8112 of ~19k message contents — every MCP tool output.

Bug 2 (found while reviewing a fix for Bug 1): ToolAnalyzer.analyze_tool_usage silently depends on Bug 1's broken output.

ToolAnalyzer.analyze_tool_usage (main_benchmark.py:458-460) detects failed tool calls by grepping recorded content for repr artifacts — "'is_error': True", "'tool_use_id': '...'" — patterns that only ever existed because the current code stores str() of the raw block list. A correct fix for Bug 1 that returns clean extracted text instead makes this grep always come up empty: failed_tool_ids stays permanently empty, and _validate_expected_tools counts every failed tool call as successful, for every task in tasks.json. A silent regression across the whole benchmark, caused by fixing Bug 1 correctly — the two need to be fixed together, not sequentially.

Also, once a proper extractor is in place, some edge cases need care: dict-shaped content still hitting a str() fallback and re-mangling codepoints; a falsy toolUseResult['result'] overwriting real content from a block's own content field; and a single root-level toolUseResult being applied to every tool_result block when a message contains more than one (e.g. parallel tool calls), rather than just the one it actually corresponds to.

I already have a fix scoped covering all of the above, plus new test coverage for SessionManager and ToolAnalyzer (currently zero coverage on this parsing path). PR to follow.