[CodesbyFebin]

Stacked on #41. Do not merge yet.

• The advisor ('Explain this round', Round Proof, mainnet review) uses the selected downloaded LiteRT-LM model (existing allowlist: Qwen2.5-1.5B-Instruct q8, 1,597,931,520 bytes, Apache-2.0, SHA-256 pinned). It is downloaded only after opt-in and is not bundled; APK size is unchanged.
• Read-only: text in, text out. No access to the wallet, MWA, the receipt signer or keys; it never edits the review, amounts or presets (AdvisorExplainerTest checks this structurally).
• Every answer starts with 'AI explanation, may be wrong; check the review.' APR/yield/profit/earnings/guarantee wording is rejected and the rule-based text is shown instead.
• Rule-based fallback when no model is downloaded, when the runtime fails (unsupported phone), or when the answer is empty.
• Options considered are in docs/on-device-ai.md (ML Kit GenAI/Gemini Nano, newer LiteRT models, llama.cpp).
• Not verified: generation on a physical phone.