fix: require exact contextual integer to float typing

LANGUAGE.md

@@ -31,6 +31,8 @@ roadmap and milestone history.
 - target-dependent C scalar primitives from `c_char` through `c_longdouble`, kept semantically distinct from native scalars
 - contextual integer/float/character literals, backward type-demand inference through names and arithmetic/bitwise expressions, and typed compile-time evaluation for numeric constant expressions
 - strict numeric conversion by default, widening where valid, C scalar coercions at C boundaries, and explicit scalar casts through keywords or transparent aliases, such as `i32(x)`, `c_float(x)`, or `StringId(x)`
+- implicit contextual integer constants must be exactly representable in the destination float, including folded expressions and inferred constants; `f32` rejects `16777217` but accepts `16777218`, while `f64` rejects `9007199254740993` but accepts `9007199254740994`; powers of two remain exact
+- float literals retain IEEE rounding, and explicit scalar casts such as `f32(16777217)` or `f64(9007199254740993)` deliberately permit rounding; existing typed-value widening and C scalar coercion rules are unchanged
 - compile-time `minval!(T)` and `maxval!(T)` bounds for concrete native, C, and scalar-backed distinct integer types; the result retains `T`
 - arrays `[N]T`, inferred-count arrays `[_]T`, sentinel arrays `[N;S]T`, compile-time expression array counts, slices `[]T` / `[;S]T`, single-item pointers `@T`, many-item pointers `*T`, and sentinel many-item pointers `[*;S]T`
 - pointer mutability via `mut`, optional pointers as nullable pointers, pointer arithmetic for many-item pointers, postfix dereference `^`, and trapping optional unwrap `?`