coef_train: un-stale the risk/lam deploy warning (libcvc #413 landed)

[transfix]

The coef_train exporter printed, for any risk/lam-widened net:

NOTE: this is a torch research artifact. The pure-C++ host's coef_feats does not build the risk-lookahead column, and its coef_mlp/drive do not read a lam output, so do NOT deploy this over share/cvc/nav/coef_mlp.cvcnav until the coordinated cvc::nav update lands.

Both claims are now false. transfix/libcvc #413 landed the coordinated cvc::nav update: coef_feats builds the risk-lookahead (7th) column, coef_mlp carries the has_mu/has_risk/has_lam flags, and drive_step_material reads the 4th output as the per-agent learned lam_soft.

This replaces the "do NOT deploy" notice with the real operational requirement:

• a libcvc host at/past #413, and
• driving through the material path (drive_step_material / the new fused drive_step_material_ext), with a material stack + grip attached — the plain drive_step still (correctly) rejects a risk net.

Also corrects the --learned-lam inline comment that repeated the stale "needs the C++ update" claim. The full round-trip (export → .cvcnav → reload → drive, incl. the learned-lam reroute) is pinned by libcvc nav_material_deploy_test. Text/comment only — no behavior change.