fix(math): powf underflow sets ERANGE regardless of exponent sign (glibc parity)

powf's range-error branch required `y > 0.0` in its underflow case, so a
negative-exponent underflow (e.g. powf(2, -200) -> 0) silently skipped
ERANGE while glibc raises it. The f64 pow already uses the sign-agnostic
`out == 0.0 && x != 0.0`; aligned powf to match.

Verified against the gcc -lm oracle (powf(2,-200), powf(10,-50),
powf(0.5,200) all ERANGE). Added two debug-mode unit tests
(powf_underflow_{negative,positive}_exponent_sets_range_errno) — the in-crate
debug harness is the one that actually exercises fl's powf body (the release
integration path const-folds/interposes constant-arg powf calls). 116 math_abi
unit tests pass; conformance_math_errno gate green.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: Dicklesworthstone

crates/frankenlibc-abi/src/math_abi.rs

@@ -1058,8 +1058,11 @@ pub unsafe extern "C" fn powf(x: f32, y: f32) -> f32 {
     if x.is_finite() && y.is_finite() {
         if x < 0.0 && y.fract() != 0.0 {
             set_domain_errno();
-        } else if out.is_infinite() || (x == 0.0 && y < 0.0) || (out == 0.0 && y > 0.0 && x != 0.0)
-        {
+        } else if out.is_infinite() || (x == 0.0 && y < 0.0) || (out == 0.0 && x != 0.0) {
+            // Range error: overflow (inf), pole at x==0 with y<0, or underflow
+            // to zero. Mirrors f64 `pow`; the underflow case applies regardless
+            // of the exponent's sign (an earlier `y > 0.0` guard wrongly skipped
+            // negative-exponent underflow, e.g. powf(2, -200), which glibc flags).
             set_range_errno();
         }
     }
@@ -8091,6 +8094,33 @@ mod tests {
         assert_eq!(abi_errno(), libc::ERANGE);
     }
 
+    #[test]
+    fn powf_underflow_negative_exponent_sets_range_errno() {
+        // Regression: powf's underflow branch used to require `y > 0.0`, which
+        // wrongly skipped negative-exponent underflow. glibc flags ERANGE for
+        // powf(2, -200) -> 0 (matches f64 pow).
+        set_errno_for_test(0);
+        let x = std::hint::black_box(2.0_f32);
+        let y = std::hint::black_box(-200.0_f32);
+        // SAFETY: ABI entrypoint accepts plain f32 input.
+        let out = unsafe { powf(x, y) };
+        std::hint::black_box(out);
+        assert_eq!(out, 0.0);
+        assert_eq!(abi_errno(), libc::ERANGE);
+    }
+
+    #[test]
+    fn powf_underflow_positive_exponent_sets_range_errno() {
+        set_errno_for_test(0);
+        let x = std::hint::black_box(0.5_f32);
+        let y = std::hint::black_box(200.0_f32);
+        // SAFETY: ABI entrypoint accepts plain f32 input.
+        let out = unsafe { powf(x, y) };
+        std::hint::black_box(out);
+        assert_eq!(out, 0.0);
+        assert_eq!(abi_errno(), libc::ERANGE);
+    }
+
     #[test]
     fn sqrt_negative_sets_domain_errno() {
         set_errno_for_test(0);