test(inference): require Rust-owned observed information

tests/test_observed_information_rust_ownership.py

@@ -0,0 +1,95 @@
+"""Fail-first ownership contracts for public observed-information arithmetic.
+
+The public inference API may validate and marshal in Python, but Hessian
+construction, symmetrization, and positive-definiteness diagnostics determine
+reported uncertainty and therefore belong to the compiled Rust numerical core.
+The tests install unmistakable compiled-core sentinels and require the public
+API to transport those exact results instead of independently reproducing the
+numerics in Python.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+
+import fast_mlsirm._core as core
+import fast_mlsirm.inference as inference
+from fast_mlsirm.config import FitConfig
+from fast_mlsirm.types import MLSIRMParams
+
+
+def _small_mirt_problem() -> tuple[np.ndarray, np.ndarray, MLSIRMParams]:
+    """Return a tiny identified-shape MIRT fixture for ownership delegation."""
+    responses = np.array([[1.0], [0.0]], dtype=np.float64)
+    factor_id = np.array([0], dtype=np.int64)
+    params = MLSIRMParams(
+        theta=np.array([[0.25], [-0.25]], dtype=np.float64),
+        alpha=np.array([0.1], dtype=np.float64),
+        b=np.array([0.0], dtype=np.float64),
+        xi=np.zeros((2, 1), dtype=np.float64),
+        zeta=np.zeros((1, 1), dtype=np.float64),
+        tau=-2.0,
+    )
+    return responses, factor_id, params
+
+
+def test_public_observed_information_delegates_matrix_to_rust(monkeypatch) -> None:
+    """The exact public Hessian result must come from the Rust numerical owner."""
+    responses, factor_id, params = _small_mirt_problem()
+    config = FitConfig(model="MIRT", backend="rust")
+    packed_dimension = int(inference._pack(params, config.normalized_model()).size)
+    sentinel = np.diag(np.arange(1, packed_dimension + 1, dtype=np.float64))
+    calls: list[tuple[tuple[object, ...], dict[str, object]]] = []
+
+    def fake_observed_information(*args, **kwargs):
+        calls.append((args, kwargs))
+        return sentinel.copy()
+
+    monkeypatch.setattr(
+        core,
+        "observed_information",
+        fake_observed_information,
+        raising=False,
+    )
+
+    result = inference.observed_information(
+        responses,
+        factor_id,
+        params,
+        config=config,
+        backend="rust",
+        device="cpu",
+        step=1e-4,
+    )
+
+    assert len(calls) == 1
+    assert np.array_equal(result, sentinel)
+
+
+def test_public_second_order_diagnostic_delegates_to_rust(monkeypatch) -> None:
+    """Positive-definiteness evidence must come from the Rust numerical owner."""
+    information = np.array([[4.0, 1.0], [1.0, 3.0]], dtype=np.float64)
+    sentinel_eigenvalues = np.array([-7.0, 42.0], dtype=np.float64)
+    calls: list[tuple[tuple[object, ...], dict[str, object]]] = []
+
+    def fake_second_order_test(*args, **kwargs):
+        calls.append((args, kwargs))
+        return {
+            "passed": False,
+            "min_eigenvalue": -7.0,
+            "eigenvalues": sentinel_eigenvalues.copy(),
+        }
+
+    monkeypatch.setattr(
+        core,
+        "second_order_test",
+        fake_second_order_test,
+        raising=False,
+    )
+
+    result = inference.second_order_test(information, tol=1e-8)
+
+    assert len(calls) == 1
+    assert result["passed"] is False
+    assert result["min_eigenvalue"] == -7.0
+    assert np.array_equal(result["eigenvalues"], sentinel_eigenvalues)