Research/ct gmarl

[elprofesoriqo]

Identified Issues

1. Red Actions Silently Ignored (Import-Time Gap)

• Symptom: Red agent moves (SCAN, EXPLOIT) return as None and do nothing.
• Cause: Missing explicit import of the netforge_rl.actions package in parallel_env.py. The global ActionRegistry remains empty during initialization.

2. Missing Centralized Physics Interface

• Symptom: Centralized critics fail because they cannot access the ground-truth state.
• Cause: The environment lacks the global_state_vector() method. Please implement it to output the 512-dim flattened state (privileges, statuses, energy).

3. Topology Initialization Crash

• Symptom: Environment crashes with AttributeError: 'NetForgeRLEnv' object has no attribute 'ordered_hosts'.
• Cause: The canonical host list is not being correctly generated from the environment's global state upon startup.

4. Replay Buffer Insertion Crash

• Symptom: A RuntimeError is thrown when inserting step data into the replay buffer, halting the training loop.
• Cause: Successor observations (next_obs) and ground-truth states are not properly sharded per agent before buffer insertion.