invalid-nan-comparison
Referencecomparison with nan is a compile-time constant
correctness · error · stable · since 1.0 · fix: never · on by default
IEEE 754 makes nan unordered with everything, itself included: =,
<, >, <=, and >= against nan are always false and <> is
always true, so the comparison never tests what it spells.
(* bad *) x = nan(* good *) Float.is_nan xFires when =, <>, <, >, <=, or >= resolves to its Stdlib
declaration with Stdlib.nan or Stdlib.Float.nan as an operand.
Shadowed operators or nan names, nan reached through an alias
(let m = nan in x = m — identity, not dataflow), compare and its kin
(a defined total order), physical equality (not constant: the same box
compares true), and other float constants deliberately do not fire. The
rewrite to Float.is_nan changes a constant answer into the intended
test — unsafe by policy. No automatic fix in this release — the promise
flips to Sometimes when that rewrite lands.

