fix: production audit gaps — retry logic + tree batching

[KJ-AIML]

Summary

Addresses critical scaling and durability gaps from PRODUCTION_AUDIT_v2.md.

Changes

SQLite Write Retry (Issue #7)

• now wraps all write paths (19 total):
• Memory CRUD: insert, delete, update layer/confidence/timestamp
• Tree: insert_node, clear_nodes, rebuild_tree_atomic
• Provenance & OTel: save, clear, prune
• Consolidation: consolidate, consolidate_and_log, consolidate_with_audit
• Triggers & expiry: log_trigger, expire_old_memories
• Removed unused decorator dead code

Tree Rebuild Batching

• now caps batch size at 20 memories
• For >20 memories, splits into batches and clusters each independently
• Prevents massive LLM prompts that fail at scale

Test Compatibility

• Added defensive calls in test fixtures
• Enables future connection-pooling work without test breakage

Test Results

• 490 tests passing, 2 skipped
• Ruff check: clean
• Format: clean

Notes

• Connection pooling was attempted but reverted due to Windows file-locking issues with singleton SQLite connections. Will be revisited with a proper connection pool architecture.
• Tree batching is LLM-path only; fallback keyword clustering was already scalable.

Closes #7