Variable: WIDTH_SAFETY_FACTOR

const WIDTH_SAFETY_FACTOR: 1.03 = 1.03

Conservative safety factors applied by the shrink solver (not by the pure measureHeightPt). Raw hmtx advances summed without shaping still pack a line a little tighter than PowerPoint, which lays text out at a device DPI and rounds each glyph advance up, so PowerPoint wraps marginally earlier and its fitted box is marginally taller. Inflating measured width and height by these factors makes the computed fontScale ≤ PowerPoint's across the calibration oracle (test/read/autofit-calibration-oracle.test.ts: 0 violations over the Aptos cases). They err on the side of shrinking slightly too much, never overflowing.