D467 - A packed half-float widens through the driver's own conversion

Status: assumed Date: 2026-09-02

A 16-bit packed float component widens to 32 bits by narrowing the field to an unsigned 16-bit value, bitcasting it to a 16-bit float type, and converting with the driver's own float-conversion instruction.

Why: a hand-rolled bit unpack gets subnormals, infinities and NaNs wrong silently unless denormals are also preserved through the pipeline, which they are not here. Routing the conversion through the driver's own IEEE path makes those cases the driver's problem instead of a constant this project would have to get right unobserved.

Rejected:

• An extended-instruction unpack: equally correct, but needs new header and type-emission infrastructure this decision does not otherwise require.
• A manual bit-unpack ("magic multiply"): needs no new infrastructure, but its subnormal path silently reads back as zero without a preserved-denormals declaration.