[RL][Phase 6] Orchestrate one AlphaZero training iteration

[utilForever]

Parent: #1058
Epic: #1052

Goal

Run self-play, replay sampling, training, and candidate creation as one reproducible iteration.

Scope

• Define the minimal shared run configuration
• Generate the configured number of self-play episodes
• Add valid samples to replay
• Train a candidate with either the C++ or Python trainer
• Preserve the current best model until evaluation succeeds
• Record seeds and artifact locations

Non-goals

• A workflow service
• Distributed workers
• Hyperparameter sweeps

Acceptance criteria

• Either trainer frontend can produce a candidate from the same replay input.
• Fixed configuration and seeds reproduce smoke-run artifacts within documented tolerances.
• A failed step leaves the current best model and replay metadata valid.
• Invalid or incomplete candidate artifacts are never evaluated.

Verification

Add one small iteration smoke test for each trainer frontend.