Just adding some things I've tried to fix this to no avail.

use RECIP_FF/CLAMPED instead of RECIP_IEEE
use RECIPSQRT_FF/CLAMPED instead of RECIPSQRT_IEEE
use MUL/MULADD/DOT4 instead of IEEE

Split constants into DOT4, rewrote DOT4 output to go via another register.

Disable scaled 16-bit float vertices, this change the misrendering to something
worse for some reason.

I've noticed the broken drawing operations are the same vbo/ib with a different
set of constant values for VS constants 124-126, the first draw works and draws
the inside gear, the second draw to draw the outside gear fails.

I considered then it might be constant caching or vertex caching since we use
vfetch to fetch the consts, tried add more flushing between draws, tried using
uncached descriptor for constant buffer.

Nothing above has fixed it, merely logging it for my own sanity at this point.