Bisecting the point where llama.cpp falls off a performance cliff on a 16GB card, why it looks like a model problem when it is a platform one, and why idle free VRAM does not predict it.

Wiring per-node error outputs into one structured log so an unattended workflow cannot fail quietly, and the trap where a workflow validates clean while its error path is unwired.

Why a read-only scout, a builder forbidden from shipping, and a separate auditor beat one general-purpose agent — and why the tool list is the real specification.