Stabilize ChaosGrad defaults and preserve trainer/optimizer API compatibility

[Copilot]

RealNet training stack needed safer defaults and backward-compatible APIs after expanding ChaosGrad and trainer configurability.

• API compatibility: Kept new max_grad_norm and ChaosGrad hyperparameters keyword-only/append-only to avoid breaking positional callers; documented new args.
• Optimizer behavior: Tuned plateau noise scaling to rely on gradient magnitude rather than near-zero params; clarified loss_history_min semantics to match retained-history behavior.
• Trainer robustness: Synced gradient persistence with AMP scaling and exposed configurable gradient clipping without shifting existing positional arguments.

Example:

optimizer = ChaosGrad(
    model.parameters(),
    lr=3e-4,
    adaptive_ema=0.2,
    plateau_noise_intensity=0.5,
    loss_history_min=64,        # retained history floor
    sentinel_threshold=1e-5,    # input health threshold
)

trainer = RealNetTrainer(
    model, data, optimizer,
    max_grad_norm=1.0  # keyword-only, preserves positional API
)

---

📱 Kick off Copilot coding agent tasks wherever you are with GitHub Mobile, available on iOS and Android.

[Copilot]

Copilot wasn't able to review any files in this pull request.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.