[alii]

decimal.to_float rounds once, in the VM, instead of turning each half into a Float first.

It used to be float.from_int(units) / float.from_int(10^scale). Both halves clamp at the largest Float, so once either passed it the answer was wrong: 10^400 at 399 places read as 1.0 instead of 10.0. Every operand of a big enough decimal hit it, and mul reaches those scales on its own.

Now a private float__from_decimal(units, scale) builds <units>e-<scale> and parses it, so Rust's own decimal-to-binary conversion rounds to the nearest Float, subnormals included, and Value::float keeps the no-NaN rule: too large is the largest Float, too small is 0.0.

Checked in the test: 10^400 at 399 places is 10.0; a negative one keeps its sign; two decimals past the largest Float both give the largest Float; 10^-400 is 0.0; and new(1, 1) is the same Float as the literal 0.1.

All seven CI gates pass locally. hawk ran for the host target, and nothing here is gated on a cfg.