export function ISERROR(value: AcceptError<Scalar>): Scalar { ... }
export function IFERROR(value: AcceptError<Scalar>, valueIfError: AcceptError<Scalar>): Scalar { ... }
// Both parameters: an unused error fallback must pass through untouched (IFERROR(5,NA()) → 5).
export function COUNT(number: AcceptError<IRange<number>>[]): number { ... }
// COUNT(NA()) is 0: the error is a value, and not a number.
The engine does not auto-reject a bare error at this parameter: it is delivered RAW — a
FormulaError.Known, never boxed as a 1×1 range — and the body narrows withFormulaError.isError(if it cares). A raw error is therefore never confusable with a one-cell range holding an error MEMBER: the two kinds this file keeps apart stay apart in the value itself, and members are no engine concern at all. For the functions whose job is to inspect, replace, or count errors — the IS* family, IFERROR, ERROR.TYPE, TYPE, COUNT and COUNTA — for the binders (LET,LAMBDA), whose value slots carry an argument unevaluated, and for a body that owns a reading ORDER or a LIFT the engine cannot see (SERIESSUM's coefficients; CONCATENATE broadcasts the error over its lifted ranges, so the SHAPE of the answer is the marker's work).On a TYPED range parameter (
IRange<number>) the marker also lets error members and failed coercions ride through the typed read as raw members — COUNT ignores them, COUNTA counts them. On an untyped range it changes nothing about members: they arrive raw either way.