Fix pickle deserialization vulnerability in RAG cache #447

Problem

src/gaia/rag/sdk.py:1620 uses pickle.load() on cached data without any validation:

cached_data = pickle.load(f)

A corrupted or tampered cache file can cause crashes or arbitrary code execution. There is no size limit check, no format validation, and no hash verification before deserialization.

Impact

• Corrupted .pkl files crash the RAG system permanently
• No integrity verification of cached data
• Potential code execution via crafted pickle files (local attack surface)

Proposed Fix

1. Add format version header (GAIA_CACHE_V1) to cache files
2. Store and verify SHA-256 checksum before unpickling
3. Validate file size before loading (reject unreasonably large caches)
4. Wrap pickle.load() in robust try/except with specific exception types (EOFError, pickle.UnpicklingError, KeyError)
5. On any load failure: delete corrupted cache file, log warning, re-index
6. Long-term: plan migration to JSON metadata + numpy embeddings

Files

• src/gaia/rag/sdk.py (lines 1615-1630)

Acceptance Criteria

• Corrupted cache files are detected and deleted automatically
• Cache load failures trigger clean re-index, not crash
• Unit test for corrupted cache recovery