[RL][Phase 4] Verify C++ and Python training parity

[utilForever]

Parent: #1056
Epic: #1052

Goal

Detect drift between the native and Python training frontends while keeping one model implementation.

Scope

• Use one fixed CPU model seed and mini-batch
• Compare forward outputs, component losses, gradients, and one simple optimizer step
• Define explicit numerical tolerances
• Test tiny-batch overfitting in both frontends

Non-goals

• Bitwise equality across GPUs or platforms
• Optimizer-state interchange
• Benchmarking frontend speed

Acceptance criteria

• Forward outputs, losses, and gradients match within documented CPU tolerances.
• One equivalent optimizer step produces matching weights within tolerance.
• Both frontends overfit the fixed batch to the agreed threshold.
• The parity test reports the first mismatched tensor clearly.

Verification

Run the cross-runtime parity test from a clean RL-enabled build.