fix: complete Uteke memory integration — inject context + extract findings

[ajianaz]

Problem

The --memory and --learn flags were added in PR #256 (issue #232) but two critical pieces are incomplete:

1. Phase 2: memory_context is recalled from Uteke but never injected into the LLM prompt. There is a TODO marker in main.rs and the variable is unused (let _ = &memory_context;)
2. Phase 3: save_findings() always sends total_issues: 0 and empty severity/categories because the actual findings are not extracted from ReviewResult

Acceptance Criteria

• Recalled memory context is appended to the LLM system prompt as additional context
• After review completes, actual findings count, severity summary, and categories are extracted from the review result and saved to Uteke
• --memory flag visibly changes review output (e.g. "Using N memories from Uteke" in stderr)
• --learn saves real data (not zeros)
• Existing tests still pass
• No clippy warnings