Address PR #391 R2 review (1 P3): tighten snapshot test to assert_array_equal

Reviewer noted the snapshot test docstring claims "byte-for-byte / bit
identity" but the assertion was assert_allclose(atol=1e-14, rtol=1e-14),
which is slightly weaker than the wording suggests.

Switch to np.testing.assert_array_equal so the assertion strength
matches the stated contract. Snapshot values are either exact f64
(Rademacher = +/-1.0) or computed once via correctly-rounded IEEE 754
sqrt in Rust (Mammen's golden-ratio constants, Webb's sqrt(N/2)
constants), so cross-platform bit-equality holds on conformant
hardware. Inline comment added explaining the safety argument.

If a runner ever fails this test on a hardware/sqrt drift basis (very
unlikely on x86_64 / aarch64 / ARM64 Windows), we'd learn something
useful about that runner before deciding to soften the check.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: igerber

tests/test_rust_backend.py

@@ -141,11 +141,13 @@ def test_bootstrap_weights_bit_identity_snapshot(self):
         }
         for weight_type, expected_arr in expected.items():
             actual = generate_bootstrap_weights_batch(2, 4, weight_type, 42)
-            np.testing.assert_allclose(
+            # Strict bit-identity: the snapshot values are either exact
+            # (Rademacher = +/-1.0) or computed once via correctly-rounded
+            # IEEE 754 sqrt in Rust (Mammen, Webb), so cross-platform
+            # bit-equality holds on conformant hardware.
+            np.testing.assert_array_equal(
                 actual,
                 expected_arr,
-                atol=1e-14,
-                rtol=1e-14,
                 err_msg=f"{weight_type} bootstrap weights drifted from pinned baseline",
             )