Cls(*xs, kw=v) — the constructor arm of the star-follower fix

What works and what does not

Fixed on 2026-09-12 (PyStarTrailingKwMinSlot + a cap in PyStarExpandCallArgs):

CPython accepts all five. test/test_nilpy_a_keyword_argument_after_a_star_unpack.npy covers the four that work and says in its own header not to add a ctor row.

Why this one was left, and it is not difficulty

The cap is already computed and already correct for this callee — the site passes a literal -1 to decline it, with the reason inline. What differs is the bookkeeping on the other side of the call:

• Every METHOD site resolves a keyword through PyKwArgIndex, which answers a 1-based parameter slot and is exactly what the cap is expressed in.
• PyClassCreate resolves a keyword to a field index (kwFld := kwPk - 1), falls back to FindUField for the dataclass shape, has a third arm for a ctor taking **kwargs (the key travels as -(node+1) on ASTIVal), and maintains nArgs, kwAny, kwExtraHead/kwExtraLast that the star expansion does not touch.

So the method fix's shape does not transfer by inspection, and the failure mode if it is wrong is the one this fix was careful to avoid: an argument that lands in the wrong slot, at run time, with no diagnostic. A refusal is the better wrong answer until someone can verify it.

How to do it

1. Add a ctor row to a differential — Ctor(*xs, forced=9) against __init__(self, a, b, forced=0), plus the dataclass shape (no __init__, keyword naming a FIELD) and the **kwargs ctor, because those are three different arms of the same loop and only the first resembles a method.
2. Pass PyStarTrailingKwMinSlot(ctorPi) instead of -1 at the site, and work out what nArgs must be afterwards — the expansion appends total - firstSlot positional args without incrementing it.
3. The positive control is that the pin, and HEAD before the change, refuse it.

What must stay refused either way

A trailing POSITIONAL (Cls(*xs, 5)) and a trailing **mapping. Those need the star's run-time length to know which slot the follower lands on, which a compile-time expansion does not have — see feature-n-a-method-call-cannot-take-an-argument-after-a-star-unpack, which keeps that half.