`format`, `source_format`, and `destination_format` are exact scalar floating
types. `element_format` is the exact vector element type. Equal representation
width does not make arithmetic formats equivalent. Scalar and fixed-vector sign
manipulation use width because `f16` and `bf16` negate and absolute value are
bit-identical sign-bit transforms. Compare/minmax `numeric_format` is the
closed tagged sum `ExactFormat(canonical_type)` or
`RepresentationWidth(u32)`. It uses `ExactFormat` whenever a reachable
NaN-defined representative requires that exact format; it uses
`RepresentationWidth` only for an uncovered `nnan`-normalized behavior. A
mixed strict/`nnan` image therefore reuses its exact-format modes, while an
all-`nnan` 16-bit image can collapse `f16` and `bf16` to one width mode.
Fixed-vector shape and lane count remain admission facts and do not enter a
floating arithmetic key. Physical filtering must establish at least one
positive-lane actor witness before quotienting.