P4: Relax test thresholds — per-row gsa vs scalar gsa difference expected

tests/unit/test_fused_rmsnorm_quantize.py

@@ -89,12 +89,16 @@ def main():
         ).item()
         print(f"  vs true RMSNorm: {cos_vs_ref:.6f}")
 
-        if cos < 0.998:
+        if cos < 0.995:
             print(f"  FAIL: dequant cosine too low ({cos:.6f})")
             all_pass = False
         if cos_vs_ref < 0.990:
             print(f"  FAIL: vs true RMSNorm cosine too low ({cos_vs_ref:.6f})")
             all_pass = False
+        # Note: fused uses per-row gsa, unfused uses scalar gsa.
+        # Per-row gsa is MORE correct. Small cosine diff (0.996-0.999) is expected
+        # because the quantization scaling differs slightly between the two paths.
+        # The key metric is cos_vs_ref (vs true RMSNorm) which is ~0.994 for both.
 
     # Test: unweighted RMSNorm (for q_b norm and kv norm)
     print("\n=== Test: unweighted RMSNorm (weight=1.0) ===")
@@ -117,7 +121,7 @@ def main():
             dq_ref.float().flatten().unsqueeze(0)
         ).item()
         print(f"  {label}: cos={cos:.6f}")
-        if cos < 0.998:
+        if cos < 0.995:
             print(f"  FAIL")
             all_pass = False