Narrowing and runtime behaviour

narrow: checked at runtime, precise statically

Narrowing a dynamically-shaped value uses a witness instead of a per-shape helper:

use_2x3(narrow(m, np.zeros((2, 3))))  # checked at runtime, precise statically

narrow returns m itself when its shape and dtype equal the witness's, and otherwise raises numty.NarrowError (a ValueError, and also an AssertionError for older callers), e.g. narrow: shape (3, 2) does not match witness (2, 3). The check is an explicit raise, not an assert, so it still runs under python -O.

isinstance and the heads

At runtime the heads are annotations, not array classes: calling F64() raises TypeError. isinstance(x, F64) is true exactly when x is an ndarray whose dtype equals native float64 (likewise every head with its own dtype, e.g. U8 and uint8; a byte-swapped '>f8' array is not an F64), so when mypy narrows a union of arrays by head, the runtime branch agrees. Rank and dims are not checked at runtime. A subscripted head (isinstance(x, F64[3])) raises TypeError like any subscripted generic, except a multi-word string spec, which erases to the bare head (isinstance(x, F64["a b"]) is the dtype check alone); use narrow or compare x.shape for dims. issubclass(np.ndarray, F64) stays false.