Seventeen commits in the window, two merged to main. Most of it went on two things that look small from outside: a phone menu 10 % too big, and a shadow pass that cost nothing.
The menus were not zoomed
The report was that phone menus look oversized and the multiplayer menu overflows. The obvious suspect was the zoom, so I printed it: on a 2364x1080 window the canvas is 1600x730 design units and the zoom clamps to 1.00, the desktop value. The menus were the desktop layout at 1.4775 px per unit, so a 76-unit button is 112 px, or 44.2 dp, half a pixel above the 44 dp thumb floor. A 10 % cut lands at 39.8 dp, so the shrink could not be a zoom.
It is one scale of 0.90 on content, on phone windows only, never on touch targets: title rows 142 to 129 px, multiplayer tiles 136 to 123 px, picker tiles 296 to 266 px tall. At 1600x900 every rect is identical.
The overflow was the title screen's multiplayer panel: content-sized but anchored at 17 % of the canvas, its foot landed at 1156 px on a 1080 px window, 76 px past the bottom edge, with the "Racing as" row clipped. After layout it is pulled into the band: foot 1045 px, 35 px of margin.
Seats that stay seats
A grandstand beside the car used to become a grey box at 200 m while its crowd was still drawing. Rows and posts are one MultiMesh per stand now, 3 draws instead of one per row plus one per 12 m, and that bought real seats out to 420 m. The structure test covers MultiMesh geometry now.
The shadow study said no
Walls are 7,680 of the 9,164 shadow primitives a chase frame submits, 84 % of them. I chunked the near wall into 40 m pieces and switched cast_shadow off past a distance: it removed 1,440 primitives for 0.00 ms, against 2.82 ms for the whole sun shadow system. The caster family is not the lever; the shadow map is. Not merged.
The ladder that would price each shadow rung never ran: the box was saturated by other work and the null control swung 46 ms, so nothing from it is reportable.
Next: the shadow rungs on a quiet machine, and the Rocket Car in a full lobby.