The float ** operator evaluates its exponent twice

Found 2026-08-16 while reading PyMakePow for [[bug-nilpy-float-overflow-answers-inf-where-cpython-raises]] — not from a report. Nothing failed; the VALUE was always right.

Repro

n = 0
def f():
    global n
    n = n + 1
    return 2.0
print(1.5 ** f())
print("calls:", n)      # CPython 1    NilPy 2

Cause

PyMakePow's static float route sends the base through pypow_dom (the guard that refuses a negative base with a fractional exponent by name) and then calls the RTL's Power. The exponent reaches both: once inside the guard's argument list, once as Power's second argument. The AST is a tree, so a node referenced twice is EMITTED twice, and CloneAST does not help — a clone is a second subtree, hence a second call.

math.pow's lowering, in the same file, already parks its exponent in a temp via PyEvalOnce and its comment says exactly why. This arm was written from that one and did not copy the line. Same shape as every other entry in [[project_nilpy_lvalue_vs_selector_path_must_both_know]]: one concept, N independent sites, and the one that stays broken is the one nobody diffed.

Fix

ppExp := PyEvalOnce(ppExp); before the guard's argument is built, in PyMakePow (compiler/pyparser.inc).

Why it can matter

Only side effects are observable — every value was already correct — so the failure mode is a counter that double-counts, a generator advanced twice, a log line printed twice. That is the same class as bug-c-a-struct-assignment-used-as-a-value-runs-its-rhs-twice, and it is the class the human-written test corpus does not catch, because tests assert values.

Gate

test/test_nilpy_pow_evaluates_operands_once.npy — six spellings (static float both ways round, base-only, augmented **=, math.pow which was already right, and a variant receiver through pypow_v), each printing the call counter, diffed against CPython. Plus the per-fix loop.