[codex] Restore bilinear feature-only policy defaults

[fei-yang-wu]

Summary

• Restore the bilinear policy default to feature-only input: F(s)z, without concatenating the raw state.
• Keep feature values detached for the policy path while allowing the posterior/latent command path to receive policy gradients.
• Add and update focused coverage for bilinear policy representation, command feature width, and offline SR pretraining behavior.

Validation

• Focused RLOpt bilinear tests passed: 8 passed.
• 1024-env Dance102 smoke matched the intended latent v0 policy input shape.
• Skynet 4096-env Dance102 job 3121342 completed successfully.