Measured 2026-09-10, compiler pascal26 at HEAD

The three rows that name the defect

var d: Double; v, w: Variant;
d := 0.0;   writeln(bits(-d));            { 8000000000000000  correct }
v := 0.0;   w := -v;  writeln(bits(w));   { 0000000000000000  WRONG   }
            w := pyneg_v(v); writeln(...); { 8000000000000000  correct }

One program, one run. pyneg_v in compiler/builtin/pylib.pas:9834 does the right thing — it tests PyVarIsFloat and writes -PyVarAsFloat(p) — and the compiler is not routing -v to it. 0.0 - 0.0 is +0.0 in IEEE 754, which is exactly the answer observed, so the lowering is a subtraction.

This is the half-wired-door shape: the door exists, it is correct, and the committed path does not use it. It does not fail by refusing; it fails by answering.

The NilPy split, which is the same fact seen from the frontend

The typed/untyped parameter pair is the control: same syntax, same value, and the only thing that differs is whether the operand is a Variant.

Why it is not prio 20

Signed zero is not decoration in the one place it is most likely to be read through a Variant. atan2 sends +0 and -0 to opposite sides of the axis — lib/rtl/math.pas's ArcTan2 has a comment about exactly that, and its own signed-zero rows are asserted — so math.atan2(-0.0, 1.0) is -0.0 in CPython and +0.0 here as soon as the argument reaches the call through a list, a tuple, a dict or an untyped parameter, which is how a Python program normally holds a number. copysign, 1/x and any sign-of-zero test inherit it too.

It stays at 35 rather than higher because a program that CARES about the sign of zero and holds the value in a Variant is rare, and no corpus site is known to hit it.

How it was found

Building test/test_nilpy_math_atan_and_atan2_bit_for_bit.npy ([[bug-b-arctan-answers-nan-above-1e300-which-is-why-math-atan2-is-still-refused]]). A row wrote math.atan(-x) while looping for x in xs, and the x = 0.0 iteration disagreed with CPython. The test now writes every sign out longhand and records the absence in its header rather than asserting a known divergence against a CPython-generated oracle.