rontolisp:narrow-float-bits

(rontolisp:narrow-float-bits src format dst &key (start 0))

Narrows a packed float array src (single-float, double-float or bfloat16, any rank) into dst, a packed (unsigned-byte 16) vector of bit patterns -- f16 or bfloat16, chosen by format (:float16 or :bfloat16) -- row-major from src's flat index 0 into dst starting at flat index start. Returns dst.

This is the bulk form of rontolisp:float16-bits/rontolisp:bfloat16-bits, the inverse of rontolisp:widen-float-bits: writing a checkpoint back out at a narrower width. Every element rounds exactly as the matching scalar primitive would -- nearest, ties to even for both formats.

A bfloat16 source hands out bit patterns rather than values: at format :bfloat16 that is a plain copy, so a tensor round-trips byte for byte.

Works on the interpreter, the JVM and both WASM backends; --no-gc has no packed float array model and refuses at compile time. A bfloat16 source is the interpreter and the JVM only, since no WASM backend carries that array width.