1.5.0

std/math works at either float width, and floats have a conversion. Both halves of one hole, reported by a consumer who had already worked around it: std/math was single precision, a generic ECS accessor is double, and nothing converted between them — DS0232 said in words that a float had no conversions at all.

• f32.nearest(v) and f64.nearest(v). One spelling per float, where an integer gets three, because an integer narrowing has three intents the compiler must not choose between and a float conversion has one: IEEE rounds to the nearest representable value, in both directions. Widening is exact and still has to be written, because LANGUAGE.md promises there is no implicit widening and a promise with an exception is one a reader has to keep a list for.
• A capability may declare a parameter or return as float, meaning f32 or f64, the same one throughout the call. The width is fixed by the first argument that has one; bare literals adopt it; a call where nothing fixes it is f32. Every std/math and std/time signature is written this way, so math.clamp(v, 0, 1) works whether v is single or double precision.
• Rounding moved from the standard library's implementations to the call site. These used to apply Math.fround to every result, which stopped being correct the moment a call could be at double precision — the code protecting precision would have been the code destroying it. The compiler now wraps an f32-resolved call and leaves an f64 one alone. A single-precision call emits the identical arithmetic it did before.
• Diagnostics point at the conversion that exists. DS0230 on two floats names f32.nearest/f64.nearest instead of the three integer spellings, which used to send a reader to DS0232 and a dead end. DS0263 on a float-to-float mismatch names the conversion and the width the call resolved to. DS0233 covers a wrong conversion method on a float; no code was renumbered and none was added.
• Refused at registration rather than at a call site: a float return with no float parameter (nothing could fix the width, so every call would quietly be f32), a decorated float such as float?, and a host type named float.

A .drs file that compiled under 1.4.0 compiles under 1.5.0, to the same JavaScript, with the same diagnostics — except that f32.clamp(v) now reports DS0233 naming nearest, where it used to report DS0232 saying no conversion existed.

For a host: implement a float capability once, in double, and do not round. If you had copied std/math's old shape and were applying Math.fround inside your own implementations, that is still correct for a signature you leave written f32, and wrong for one you change to float.