fix: curator JSON parsing and CLI response extraction

[jespo2021]

Summary

Fixes memory curation failing silently due to JSON parsing errors and incorrect CLI response extraction.

Changes

• Fix JSON syntax in curation prompt: Changed example JSON from Python dict syntax (single quotes) to valid JSON (double quotes)
• Fix CLI response extraction: Handle Claude Code's nested message format message["message"]["content"] instead of message["content"]
• Add fallback JSON parsing: Use ast.literal_eval when LLM returns Python dict syntax despite instructions
• Implement real stats: /memory/stats now returns actual counts instead of placeholder zeros

Root Cause

The curator was returning 0 memories because:

1. Prompt examples used Python syntax: The curation prompt showed 'key': 'value' examples, so Claude followed that format and returned invalid JSON
2. Wrong extraction path: Claude Code CLI outputs {"type":"assistant","message":{"content":[...]}} but the code looked for message["content"] directly
3. No error recovery: When JSON parsing failed, it silently returned empty results

Testing

Regression tests added to verify the fixes:

• test_cli_response_extraction.py: Validates nested Claude Code message format handling
• test_json_parsing_fallback.py: Validates ast.literal_eval fallback for single-quote JSON
• test_storage_stats.py: Validates get_stats() returns real database counts

All tests verified to fail on main branch and pass on this branch.

# Run tests
uv run pytest tests/ -v

# Before fix (on main)
3 failed

# After fix (on this branch)
3 passed

🤖 Generated with Claude Code

[jespo2021]

I had an issue while testing on my Mac and hope this helps