From 599e9e8efe1bf014b9e04e4de72e5bf378db8356 Mon Sep 17 00:00:00 2001
From: Trent Nelson <trentn@nvidia.com>
Date: Sat, 3 Oct 2026 23:58:04 -0700
Subject: [PATCH] Avoid copying packed scientific evidence before encoding

Pass the canonical host buffer directly to SHA-256 and Base64.
Retain the exact receipt format while avoiding a full-size byte copy.

Signed-off-by: Trent Nelson <trentn@nvidia.com>
---
 src/cuphoton/xscan/device_pipeline.py |  8 +--
 tests/xscan/test_device_pipeline.py   | 77 +++++++++++++++++++++++++++
 2 files changed, 81 insertions(+), 4 deletions(-)

diff --git a/src/cuphoton/xscan/device_pipeline.py b/src/cuphoton/xscan/device_pipeline.py
index f2b7b829..0187ce43 100644
--- a/src/cuphoton/xscan/device_pipeline.py
+++ b/src/cuphoton/xscan/device_pipeline.py
@@ -2361,16 +2361,16 @@ def _scientific_evidence_receipt(
     if int(host_values.size) != expected_count:
         raise RuntimeError("terminal scientific evidence size changed")
     canonical = np.ascontiguousarray(host_values, dtype=np.dtype("<f8"))
-    packed_bytes = canonical.tobytes(order="C")
+    packed_buffer = memoryview(canonical).cast("B")
     return DevicePipelineScientificEvidence(
         candidate_count=evidence.candidate_count,
         parameter_names=evidence.parameter_names,
         feature_names=evidence.feature_names,
         layout=layout,
         packed_element_count=int(canonical.size),
-        packed_byte_count=len(packed_bytes),
-        packed_sha256=hashlib.sha256(packed_bytes).hexdigest(),
-        packed_base64=base64.b64encode(packed_bytes).decode("ascii"),
+        packed_byte_count=len(packed_buffer),
+        packed_sha256=hashlib.sha256(packed_buffer).hexdigest(),
+        packed_base64=base64.b64encode(packed_buffer).decode("ascii"),
         xfit_schema=evidence.xfit_schema,
         xfit_backend=evidence.xfit_backend,
         xfit_solver=evidence.xfit_solver,
diff --git a/tests/xscan/test_device_pipeline.py b/tests/xscan/test_device_pipeline.py
index a4f41f7a..e5a5d487 100644
--- a/tests/xscan/test_device_pipeline.py
+++ b/tests/xscan/test_device_pipeline.py
@@ -564,6 +564,83 @@ def test_device_pipeline_import_keeps_cuda_dependencies_lazy() -> None:
     assert result.returncode == 0
 
 
+@pytest.mark.parametrize("strided", [False, True])
+def test_scientific_evidence_receipt_preserves_packed_bytes(
+    strided: bool,
+) -> None:
+    import cuphoton.xscan.device_pipeline as pipeline
+
+    expected = _scientific_evidence_payload()
+    raw = bytearray(base64.b64decode(expected["packed_base64"]))
+    # Encoding must retain signed zeros, infinities and distinct NaN payloads.
+    bits = (
+        0x0000000000000000,
+        0x8000000000000000,
+        0x7FF0000000000000,
+        0xFFF0000000000000,
+        0x7FF8000000000001,
+        0xFFF8000000000002,
+    )
+    raw[: 8 * len(bits)] = b"".join(
+        value.to_bytes(8, "little") for value in bits
+    )
+    expected_bytes = bytes(raw)
+    values = np.frombuffer(expected_bytes, dtype="<f8").astype(np.float64)
+    backing = np.full(values.size * (2 if strided else 1), 12345.0)
+    host_values = backing[::2] if strided else backing
+    host_values[:] = values
+    before = backing.tobytes()
+    host_values.flags.writeable = False
+
+    layout = tuple(
+        pipeline.DevicePipelineEvidenceSegment(
+            **{**segment, "shape": tuple(segment["shape"])}
+        )
+        for segment in expected["layout"]
+    )
+    metadata = {
+        f"{component}_{name}": value
+        for component in ("xfit", "xpois")
+        for name, value in expected[component].items()
+    }
+    for name in ("status_code_names", "uncertainty_reason_code_names"):
+        metadata[f"xfit_{name}"] = tuple(
+            (int(code), label)
+            for code, label in expected["xfit"][name].items()
+        )
+    metadata["xpois_kernel_shape"] = tuple(expected["xpois"]["kernel_shape"])
+    metadata["xpois_basis_terms"] = tuple(
+        (
+            term["component_index"],
+            term["sigma"],
+            term["poly_u_degree"],
+            term["poly_v_degree"],
+            term["zero_sum"],
+        )
+        for term in expected["xpois"]["basis_terms"]
+    )
+    evidence = pipeline._PackedScientificEvidence(
+        values=None,
+        layout=layout,
+        candidate_count=expected["candidate_count"],
+        parameter_names=tuple(expected["parameter_names"]),
+        feature_names=tuple(expected["feature_names"]),
+        **metadata,
+    )
+    actual = pipeline._scientific_evidence_receipt(
+        evidence, layout=layout, host_values=host_values
+    ).to_payload()
+
+    expected["packed_sha256"] = hashlib.sha256(expected_bytes).hexdigest()
+    expected["packed_base64"] = base64.b64encode(expected_bytes).decode(
+        "ascii"
+    )
+    assert actual == expected
+    assert base64.b64decode(actual["packed_base64"]) == expected_bytes
+    assert backing.tobytes() == before
+    assert not host_values.flags.writeable
+
+
 def test_scientific_evidence_v1_roundtrip_preserves_expected_nans() -> None:
     payload = _scientific_evidence_payload()