fix: strip markdown code fences from LLM judge responses

[cwilson613]

Summary

• Anthropic models (via litellm) wrap JSON responses in ```json ... ``` code blocks even when response_format=json_object is set
• This causes json.loads() to fail, scoring every evaluation as 0/100
• Fixed in both runner.py (benchmark scoring) and asset_generator.py (eval generation)

What changed

Added _strip_code_fences() helper in runner.py and inline fence stripping in asset_generator.py. Both extract JSON from markdown code blocks before parsing. Normal JSON responses (without fences) pass through unchanged.

How I found this

Running EvalMonkey with EVAL_MODEL=anthropic/claude-haiku-4-5 against a custom agent. Every benchmark scored 0/100 with the error Expecting value: line 1 column 1 (char 0). The raw LLM response was:

```json
{"score": 90, "reasoning": "..."}

extra text here

After this fix, benchmarks score correctly.

## Test plan

- [x] All 46 existing tests pass (`python -m pytest tests/ -v`)
- [x] Verified with Anthropic claude-haiku-4-5 as eval judge — scores now match expected behavior
- [x] Normal JSON responses (OpenAI, no fences) still parse correctly

🤖 Generated with [Claude Code](https://claude.com/claude-code)

[himmi-01]

Thanks @cwilson613 .

May be in future we can remove the duplication of.

if content and "```" in content:
                match = re.search(r"```(?:json)?\s*\n?(.*?)\n?\s*```", content, re.DOTALL)
                if match:
                    content = match.group(1).strip()