Fix RL simulation boundary handling - #51

What changed

• make empty-action states terminate explicitly without mutating the controller
• encode arbitrary network node IDs as dense action/tensor indices
• handle isolated virtual nodes and empty edge tensors safely
• correct feature normalization and virtual-node progress semantics in feature schema v2
• retain automatic feature schema v1 compatibility for legacy checkpoints
• make checkpoint loading strict and versioned
• repair sample, beam, and recoverable decoding paths
• reject unsupported special/composite action configurations clearly
• use zero bootstrap values for terminal rollouts
• add focused RL boundary and searcher regression tests

Why

Several RL paths relied on contiguous IDs, non-empty action masks, non-empty graphs, or silently accepted incompatible checkpoints/configurations. Those assumptions could produce random masked actions, NaNs, invalid tensor indices, or misleading evaluation behavior.

Compatibility

Legacy checkpoints without schema metadata automatically select feature schema v1. New runs use schema v2 and save the schema version in checkpoint metadata.

Validation

• python -m pytest -q tests/solver/learning — 25 passed
• python -m pytest -q — 229 passed
• python -m compileall -q virne tests
• git diff --check
• historical PPO checkpoint smoke test on 10 corrected-v2 requests — 10/10 accepted, zero total violation