[RL][Phase 6] Validate and document the complete AlphaZero MVP

[utilForever]

Parent: #1058
Epic: #1052

Goal

Prove and document the complete fixed-deck mirror-match MVP.

Scope

• Run self-play → replay → training → arena → promotion end to end
• Run the agreed uniform-policy MCTS baseline evaluation
• Document native C++ and Python training entry points
• Document build requirements, configuration, artifacts, and current limitations
• Capture profiling data only to identify the next justified issue

Non-goals

• Fixing unmeasured performance limits
• Full-card-pool claims
• Distributed or multi-GPU instructions

Acceptance criteria

• The end-to-end smoke loop completes with either trainer frontend.
• The trained candidate scores at least 55% over 200 paired-seat games against uniform-policy MCTS using the same search budget, counting a draw as 0.5.
• The documented commands reproduce the smoke workflow from a clean supported setup.
• Default non-RL build and test commands remain unchanged.
• Any proposed optimization cites measured profiling evidence.

Verification

Run the complete focused C++ and Python test sets, both trainer smoke loops, and the 200-game paired-seat evaluation.