**= to a SUBSCRIPT target does not parse

• Type: bug (ordinary Python refused, at COMPILE time) — Track N
• Found: 2026-08-07, grepping the bug class while fixing [[bug-nilpy-truediv-and-pow-assign-on-a-class-instance-skip-the-dunder]].
• Pre-existing — identical on stable_linux_amd64/default/pinned.

Measured

xs = [2, 3]
xs[0] **= 5
print(xs)          # CPython [32, 3]

pascal26:2: error: expected expression
  near:   xs    >>>

xs[1] /= 3 and xs[0] += 1 on the same target are fine, so it is **= specifically, and only on a subscript.

Why it is a third arm

Augmented assignment has three target shapes in this frontend, each with its own code:

The common cause across all three is that ** is not a binary TOKEN — it lowers through PyMakePow, not AN_BINOP — so every arm keyed on PyAugBinTok has nothing to return for it and must special-case tkPowEq explicitly. Two of the three now do.

Fix shape

Give ParseLValueAST's augmented path a tkPowEq arm that desugars to target = target ** rhs through PyMakePow, with the target subtree handled the way augRead already handles it (the index expression must be evaluated once — d[f()] **= 2 calls f once, which is exactly why that path builds an augRead rather than duplicating the subtree).

PyAugDunderName already answers __ipow__/__pow__ for tkPowEq, so the user-class case comes free once the arm exists — but note that a subscript target holding a class instance goes through PyAugClassDunderNode, which is in pyparser.inc; check whether ParseLValueAST can reach it before assuming.

Gate

Per-fix loop. Extend test/test_nilpy_truediv_pow_assign_class_dunder.npy — which already carries the other two target shapes and marks this one as the known gap — with the subscript form over an int, a float and a class instance, diffed against CPython.

2026-08-07 — FIXED

The third arm landed exactly where the ticket sized it: ParseLValueAST's default-property subscript route in parser.inc. PyAugBinTok(tkPowEq) is tkEOF, so the route's isAssign peek never fired and the statement died on "expected expression" before any augmented handling ran. Added an augPow flag beside augBin: it is set when the token past ] is tkPowEq, drives the same augRead construction (so the read half is built over a CLONE of the index chain, as before), and combines through PyMakePow instead of AN_BINOP — which carries the __pow__ / __rpow__ dispatch with it for free.

The sibling __getitem__/__setitem__ route beside it also keyed its named refusal on PyAugBinTok(...) <> tkEOF, so obj[k] **= v on a user class fell past it into the same bare "expected expression". It now names itself.

Measured (self-hosted at HEAD + this change), diffed against CPython

Deliberately NOT fixed here

• A subscript holding a class INSTANCE (ps[0] **= 7 with __pow__) still raises TypeError: expected a number, got object. Measured the sibling: += with __add__ and *= with __mul__ fail identically, so this is not specific to **= — a container erases the element's static class and no dunder dispatch happens. That is the blocked [[bug-nilpy-dunders-not-dispatched-through-containers]] family. Parity with the other augmented operators is the bar this ticket set, and it is met.
• The index is still evaluated twice (xs[k()] **= 1 calls k twice) — the ticket's fix-shape note assumed augRead already evaluated it once; it does not, and neither does any other augmented operator on this route. That is [[bug-nilpy-augmented-subscript-evaluates-its-index-twice]], pre-existing and filed.

Test

test/test_nilpy_truediv_pow_assign_class_dunder.npy extended as the Gate line asked (it already carried the other two target shapes and a NOTE marking this one as the known gap; the NOTE is replaced by the real coverage). Output is byte-identical to CPython's for the whole file. Makefile expectation updated.

Log

• 2026-08-07 — resolved, commit 06062c8e2.