How accurate are this adapter's transcendentals?

WGSL permits relaxed precision on pow, sin and cos: an implementation may answer to a few ULP rather than correctly rounding. The rotary embedding is the only part of the forward pass that calls them, and its kernel comparison failed at 2.2e-5 against a 1e-5 bound. This measures each function separately, over exactly the angles the rotary embedding uses, so the failure can be attributed instead of guessed at.

running...