[RL][Phase 6] Promote models safely and record reproducible run metadata

[utilForever]

Parent: #1058
Epic: #1052

Goal

Promote only candidates that pass evaluation and retain enough metadata to reproduce the decision.

Scope

• Define the arena promotion threshold
• Atomically replace or reference the current best weight artifact
• Retain candidate and incumbent identifiers
• Record schema versions, seeds, search settings, trainer frontend, losses, and arena result
• Reject incompatible model or schema versions

Non-goals

• A model registry service
• Remote artifact storage
• Cross-runtime optimizer-state promotion

Acceptance criteria

• Failed candidates leave the incumbent artifact unchanged.
• Successful promotion is atomic from the consumer's perspective.
• C++ and Python load the promoted weights.
• Promotion metadata identifies every input needed to repeat the arena decision.
• Interrupted promotion cannot leave an incomplete best-model artifact.

Verification

Add focused pass, fail, incompatible-version, and interrupted-write tests.