Goal 7 — Rewrite-step prediction (gate G7)

[duckyquang]

The generative skill: from a state graph, predict (rule id, application site).

Sub-goals

• 7.0 Step dataset derived from 6.1 traces; per-rule stratification; site-labeling scheme on nodes
• 7.1 Policy head on the shared encoders (per-node site scoring + rule classification); transformer step-proposer baseline (state as prefix string → rule token)
• 7.2 Metrics: top-k verifier-valid step accuracy, rule coverage, unseen-family generalization
• 7.3 Grid across the same channels; analysis: does homogeneous topology help rule application transfer across contexts? (the sharpest EML test)
• 7R Reserved repair pass

Gate G7

Learned policy beats the uniform-random valid-step baseline by a wide margin; per-rule coverage has no dead rules.

Depends on Goal 6 (traces + encoders).