One signature, every rank

The headline: rank-generic helpers with no overloads, no per-rank copies. mypy solves *Ts from each call site, and the body checks once, rank-generically:

from typing import Literal
from typing_extensions import TypeVarTuple, Unpack
from numty import FlexArray

Ts = TypeVarTuple("Ts")

def squared_norm(
    v: FlexArray[tuple[Unpack[Ts], Literal[2]], np.dtype[np.float64]],  # type: ignore[type-var]
) -> FlexArray[tuple[Unpack[Ts]], np.dtype[np.float64]]:  # type: ignore[type-var]
    return v[..., 0] ** 2 + v[..., 1] ** 2

squared_norm(np.zeros((5, 7, 2)))  # (5, 7)
squared_norm(np.zeros((5, 2)))     # (5,)
squared_norm(np.zeros((5, 7, 3)))  # error: expected (..., 2), got (..., 3)

Two honest type: ignore[type-var] notes per signature (NumPy's stubs bind the shape to tuple[int, ...], which an unsolved *Ts cannot satisfy) buy all of this: inside, the solver threads *Ts through ellipsis indexing (v[..., 0] on (*Ts, 2) is (*Ts,), and out-of-range indices still error against the known suffix), same-*Ts broadcasts, unary ufuncs, astype/asarray, .shape, &/| on bool pairs, where/clip, comparisons, reductions along axes that resolve to a known dim (.all(axis=-1) on (*Ts, 2) is (*Ts,)), and rank arithmetic over the unknown middle — matmul on (*Ts, N, M) pairs (inner dims must align, batches broadcast through the unpack) and leading None/slice indexing without an ellipsis (v[0] on (5, *Ts, 7) is (*Ts, 7)). reshape over *Ts returns the spelled target as-is ((*Ts, 2, 3) to (6,)), and transpose/.T over *Ts bails out instead of claiming an order no annotation can spell. Anything that would need the unknown length or order — several array indices or an integer beside one (numpy may move their block to the front), an index reaching into *Ts itself, reducing over *Ts itself, mixing different Ts declarations, explicit transpose axes, a -1 that would absorb *Ts — bails out instead of guessing. A lone integer array or boolean mask on the known prefix gathers in place (frames[idx] on (total, *Ts) is (K, *Ts)).