Circuit induction_kchain_weights (GPT-2)

K-composition (weight + path-patch, GPT-2) — scope: gpt2

The induction macro read from the weights (K-composition) and path-patch-gated on GPT-2 (the cross-model behavioural view is on the induction page).

• canonical writer 4.11; 4/5 canonical edges live
• K-composition static 0.06873766100185522 vs random 0.03909262899650067; top edge rel-drop 0.5611611045403996

Data: runs/disassembly/circuits/atlas_summary.json + the discovery artifacts. Regenerate: circuit_catalog_doc.py.