Decision: call AI Gateway with plain fetch, NOT [[DESIGN-scribble-works|scribble-works]] authoringFetch, because authoringFetch reserves against SW production D1 budget ledger (brief: don't touch SW production).
Quiver balance endpoint 403 (key lacks billing_read) -> Quiver cost = provider-reported tokens x list price ($4/M in, $20/M out). Grok cost = provider usage.cost_in_usd_ticks ($0.02 each).
Prompt STYLE asked for "simple gentle faces" -> both models put faces on objects (G4 wreath, G5 church/tree, Q4 candles). Regens added "only people and animals have faces".
G1 a1: Joseph drawn as a second Mary, no ox -> regen a2 pass.
Arrow (both generation and vectorization) draws in an internal 80x80 square, letterboxed in the requested 765x990 viewBox -> art occupies a band. fit.py re-sets viewBox to inked bbox + 3% pad, applied to all 10 pages equally.
G3 vectorization emitted a 60x60 clipPath that cut the top of the star -> --unclip for G3 only (other clips matched the image bounds).
G4 first vectorization dropped wreath outline + broke a candle edge (open regions) -> re-vectorized once (a2v2), mostly closed.
Headless Chrome --screenshot writes the PNG but does not exit -> render.py polls for a stable file and kills Chrome.
Stopped generating at ~$3.12 spent; no further regens needed.