[dev] Megatron Lite (2/4) runtime layering guard

[ISEEKYAN]

Summary

Part 2/4 of the Megatron Lite foundation series. Establishes and enforces the layering contract across the bench / VERL-connector / runtime / model / primitive layers, so that model-specific metadata cannot leak into model-agnostic layers.

Motivation

packed_seq_params / position_ids are transient THD metadata. They may be materialized only at the immediate forward boundary — inside an explicitly marked allow-range in the bridge runtime — and must never be carried as model-agnostic batch fields or leaked through the connector/runtime layers. This PR makes that boundary explicit and guards it.

What's included

• Runtime backends runtime/backends/{bridge, mbridge, mlite} and runtime/megatron_utils.
• bench and VERL connector layers reworked to hand the runtime only a model-agnostic batch.
• Static layering guard: tests/unit/runtime/test_layering_contracts.py, test_packed_batch_bridge.py, tests/run_layering_contracts.sh — directional import boundaries (denylist + vetted, documented allowlist) plus the transient-metadata allow-range check.
• Optimizer-config builders renamed build_mc_* → build_dist_opt_*, with the bridge/optimizer consumers updated atomically in this PR.

Depends on

The 1/4 "shared primitive core" PR in this series.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.