Repro

x = 1024
a = x + 1
print(~a)            # pxx 4        CPython -1026
print(not a)         # pxx True     CPython False
if not a:
    print("branch: falsy")     # pxx takes THIS one
else:
    print("branch: truthy")    # CPython takes this one

What varies, measured

+, -, * and >> all trigger it, so it is not the operator. A name that has EVER been assigned from a binop stays broken even after a later assignment from a bare literal, which points at the name's inferred type rather than at the value flowing through it.

0x3000000000000004 and 0x7000000000000004 are a tag in the high nibble over a small payload — so ~ is complementing a BOX (variant / promo-int handle), not the integer inside it. ~a = 4 in the first row is the same shape with the tag cleared: whatever it reads, it is not 1025.

not returning True for every one of them is consistent with reading the same non-integer and finding it falsy — and not 128 and not y are both correct, so the truthiness path itself is fine.

Why prio 70

It is silent and it moves control flow. if not count: and if not (mask >> k): are ordinary Python; both take the wrong branch here with nothing printed. CPython compatibility is one-directional (nilpy-semantics-divergences.md) and this is the direction that is a defect: working CPython code runs wrong.

Suspected site

pyparser.inc's opIsArith decision (around the notOp := Integer(ASTIVal[left]) block) chooses bitwise-vs-logical not from the operator id of the operand node. When the operand is a NAME rather than a binop, ASTIVal[left] is not an operator id at all — it is whatever that node kind keeps there — so the branch is taken on a field that means something else. That is a guess from reading and it is the thing to disprove first; the boxed-value tags above may point somewhere lower.

Found

2026-08-31, while verifying that the tkIdent -> tkShrLogical rename (bug-a-shr-reaches-the-ir-spelled-as-tkident) had not disturbed that opIsArith line. It had not — the pinned compiler produces byte-identical wrong answers, so this predates the rename entirely. It is unrelated to the shift work except that a >> probe is what put it on screen.