A CANONICAL ANSWER · FROM THE VINDEX KNOWLEDGE GRAPH
The obvious protection made it worse
Recorded, and worth sitting with: keeping every down_proj at BF16 — the intuitive protection — cost over a gigabyte and moved granite's tail from 4.6224 to 4.8010. Worse. Keeping the last five FFN layers instead cost 431 MiB and moved it to 1.2826 — three and a half times better. Measurements can justify a map; VINDEX3 can express, compile and execute it; automatically discovering the right one remains open — four cheap screens have been tried against ground truth, and all four failed.
SUPPORTED · recorded — granite-4.1-3b: all-down_proj 4.8010 · late-5 FFN 1.2826