Record arXiv v1.1 submission

README.md

@@ -66,11 +66,11 @@ Articles and original figures use **CC BY 4.0**. Their version, SHA-256 digests
 and verification scope are in the [publication index](docs/publications/README.md).
 These preprints preserve negative results, missing measurements and the boundary
 between frozen v1 evidence and the separately labelled post-freeze v2 handoff.
-The paper uses arXiv **submission number 8127259**. On **2026-09-25**, the v1.1
-PDF and metadata were processed successfully and left at the final **Submit**
-confirmation boundary; editing the previous on-hold submission returned it to
-**incomplete** status. This tracking number is not a public arXiv article
-identifier. No announcement or peer-review acceptance is claimed.
+The paper uses arXiv **submission number 8127259**. The v1.1 PDF and metadata
+were processed successfully on **2026-09-25** and submitted on **2026-09-26**.
+The arXiv account currently reports **on hold** for moderation. This tracking
+number is not a public arXiv article identifier. No public announcement or
+peer-review acceptance is claimed.
 
 Chrome Web Store version **0.1.0** passed review and is publicly available as
 [Vons — Local Decisions](https://chromewebstore.google.com/detail/vons-%E2%80%94-local-decisions/cbmjkokoojinfgafahncmjlfmicinide),

docs/HUGGING_FACE.md

@@ -66,11 +66,11 @@ Articles and original figures use **CC BY 4.0**. Their version, SHA-256 digests
 and verification scope are in the [publication index](../docs/publications/README.md).
 These preprints preserve negative results, missing measurements and the boundary
 between frozen v1 evidence and the separately labelled post-freeze v2 handoff.
-The paper uses arXiv **submission number 8127259**. On **2026-09-25**, the v1.1
-PDF and metadata were processed successfully and left at the final **Submit**
-confirmation boundary; editing the previous on-hold submission returned it to
-**incomplete** status. This tracking number is not a public arXiv article
-identifier. No announcement or peer-review acceptance is claimed.
+The paper uses arXiv **submission number 8127259**. The v1.1 PDF and metadata
+were processed successfully on **2026-09-25** and submitted on **2026-09-26**.
+The arXiv account currently reports **on hold** for moderation. This tracking
+number is not a public arXiv article identifier. No public announcement or
+peer-review acceptance is claimed.
 
 Chrome Web Store version **0.1.0** passed review and is publicly available as
 [Vons — Local Decisions](https://chromewebstore.google.com/detail/vons-%E2%80%94-local-decisions/cbmjkokoojinfgafahncmjlfmicinide),

docs/publications/ARXIV.md

@@ -5,15 +5,15 @@ release, alongside the [v1.1 Technical Report](vons-tech-report-v1.1.pdf). The
 paper uses **submission number 8127259**. On **2026-09-25**, the v1.1 PDF replaced
 the prior v1.0 file, arXiv reported **Submission processed OK**, and the title,
 author, abstract, comments, categories and CC BY 4.0 selection were reviewed.
-The processed PDF was opened and the workflow was left at the final **Submit**
-button for explicit action-time approval. Editing the prior on-hold submission
-returned the account status to **incomplete**; the final submission has not been
-made. This is a submission tracking number, not a public arXiv article
-identifier. No public article URL or acceptance is announced. arXiv performs
-content moderation, not journal peer review; see its
+After the processed PDF was opened and explicitly approved, the article was
+submitted on **2026-09-26**. The account first reported **Article submitted**
+with status **processing**, then changed to **on hold** for moderation. This is
+a submission tracking number, not a public arXiv article identifier. No public
+article URL or acceptance is announced. arXiv performs content moderation, not
+journal peer review; see its
 [moderation policy](https://info.arxiv.org/help/moderation/index.html).
 
-## Prepared replacement artifact
+## Submitted replacement artifact
 
 | Field | Submitted value |
 | --- | --- |
@@ -24,9 +24,9 @@ content moderation, not journal peer review; see its
 | License | CC BY 4.0 |
 | PDF | `vons-paper-v1.1.pdf`, 21 pages |
 | SHA-256 | `8e9ff7d768979411db2333f37111692b5b8426bdaec21d35954c7038d047f321` |
-| Current account status | `incomplete`, awaiting final Submit confirmation |
+| Current account status | `on hold`, awaiting arXiv moderation |
 
-The prepared abstract is an abridged version preserving the frozen negative
+The submitted abstract is an abridged version preserving the frozen negative
 results, separately labelled post-freeze remediation results and manuscript
 limitations. Comments record 21 pages, 2 figures and the public Hugging Face
 source/Tech Report location. Journal reference and DOI remain blank. The Tech
@@ -34,9 +34,10 @@ Report is available alongside the paper in this repository; it is not prepared
 as a second arXiv paper.
 
 The prior v1.0 package reached on-hold status before this replacement was
-prepared. Follow any instructions sent by arXiv and do not create a duplicate
-submission. Add a public article link only after arXiv assigns and announces
-one.
+prepared. The submitted v1.1 replacement is now on hold under the same
+submission number. Follow any instructions sent by arXiv and do not create a
+duplicate submission. Add a public article link only after arXiv assigns and
+announces one.
 
 Exact local PDF hashes match the reviewed final manuscripts, and all 50 v1.1
 pages have been inspected. The documents were generated with Typst, not TeX/LaTeX.

docs/publications/README.md

@@ -52,10 +52,9 @@ original figures. Software rights are governed separately by the
 [Vons Community and Commercial License 1.0](../../LICENSE); article licensing
 does not grant commercial software rights.
 
-The paper uses arXiv **submission number 8127259**. On **2026-09-25**, the v1.1
-PDF and metadata were processed successfully and left at the final **Submit**
-confirmation boundary; editing the previous on-hold submission returned it to
-**incomplete** status. This tracking number is not a public arXiv article
-identifier. No public arXiv URL, DOI, announcement or peer-review acceptance is
-claimed. See the
+The paper uses arXiv **submission number 8127259**. The v1.1 PDF and metadata
+were processed successfully on **2026-09-25** and submitted on **2026-09-26**.
+The arXiv account currently reports **on hold** for moderation. This tracking
+number is not a public arXiv article identifier. No public arXiv URL, DOI,
+announcement or peer-review acceptance is claimed. See the
 [arXiv submission record](ARXIV.md).

tests/test_onnx_export.py

@@ -226,60 +226,113 @@ def test_pooled_diffusion_manifest_keeps_static_options(tmp_path: Path) -> None:
 
 
 def test_shape_helper_handles_dynamic_and_static_dimensions() -> None:
-    class _Dim:
-        def __init__(self, *, value=None, param=None):
-            self._value = value
-            self._param = param
-
-        def HasField(self, name):
-            if name == "dim_value":
-                return self._value is not None
-            if name == "dim_param":
-                return self._param is not None
-            return False
-
-        @property
-        def dim_value(self):
-            return self._value
-
-        @property
-        def dim_param(self):
-            return self._param
-
-    class _Shape:
-        def __init__(self, dims):
-            self.dim = [_Dim(value=d) if isinstance(d, int) else _Dim(param=d) for d in dims]
-
-        def HasField(self, name):
-            return name == "dim"
-
-    class _TensorType:
-        def __init__(self, dims):
-            self.shape = _Shape(dims)
-            self.elem_type = 1
-
-        def HasField(self, name):
-            return name == "shape"
-
-    class _ValueInfo:
-        def __init__(self, name, dims):
-            self.name = name
-
-            class _Type:
-                def __init__(self, dims):
-                    self.tensor_type = _TensorType(dims)
-
-            self.type = _Type(dims)
+    """_shape decodes protobuf ``HasField``-distinguished ``dim_value``/
+    ``dim_param`` dimensions (including the tensor_type.HasField("shape")
+    precondition).  Uses real onnx protobuf structs so the HasField gate
+    matches what bundle.py actually reads at runtime."""
+    onnx = pytest.importorskip("onnx")
+    ids = onnx.helper.make_tensor_value_info(
+        "ids",
+        onnx.TensorProto.INT64,
+        [onnx.helper.make_tensor_dimension(1),
+         onnx.helper.make_tensor_dimension("options"),
+         onnx.helper.make_tensor_dimension("sequence_length")],
+    )
+    mask = onnx.helper.make_tensor_value_info(
+        "mask",
+        onnx.TensorProto.INT64,
+        [4, 32, 512],
+    )
+    infos = _value_infos([ids, mask], onnx)
+    assert infos[0]["shape"] == [1, "options", "sequence_length"]
+    assert infos[1]["shape"] == [4, 32, 512]
+    assert infos[0]["name"] == "ids"
+    assert infos[0]["dtype"] == "INT64"
+
+
+def test_onnxruntime_variable_sequence_width_64_and_128_yield_finite_equivalent_outputs(
+    tmp_path: Path,
+) -> None:
+    """Regression: an actual onnxruntime InferenceSession with a causal,
+    attention-masked op produces finite, output-equivalent results when
+    the same 4 live tokens are wrapped in a 64-wide vs 128-wide dynamic
+    sequence container.
+
+    Uses real ort + real onnx protobuf, no stubs: build a tiny graph that
+    computes ``sum(attention_mask * input_ids, axis=-1)``, which is a
+    strict function of the live prefix only. Feeding the identical 4 live
+    tokens at widths 64 vs 128 must produce (a) finite outputs and (b)
+    byte-identical scores because the zero-masked positions contribute
+    nothing (the canonical attention-masked equivalence used downstream).
+    """
+    onnx = pytest.importorskip("onnx")
+    ort = pytest.importorskip("onnxruntime")
 
-    class _FakeOnnx:
-        TensorProto = type("TensorProto", (), {"DataType": type("DataType", (), {"Name": staticmethod(lambda x: f"type_{x}")})})
+    inputs = [
+        onnx.helper.make_tensor_value_info(
+            "input_ids",
+            onnx.TensorProto.INT64,
+            [onnx.helper.make_tensor_dimension(1),
+             onnx.helper.make_tensor_dimension("options"),
+             onnx.helper.make_tensor_dimension("sequence_length")],
+        ),
+        onnx.helper.make_tensor_value_info(
+            "attention_mask",
+            onnx.TensorProto.INT64,
+            [onnx.helper.make_tensor_dimension(1),
+             onnx.helper.make_tensor_dimension("options"),
+             onnx.helper.make_tensor_dimension("sequence_length")],
+        ),
+    ]
+    outputs = [onnx.helper.make_tensor_value_info("scores", onnx.TensorProto.INT64, [1, "options"])]
+
+    cast_ids = onnx.helper.make_node("Cast", ["input_ids"], ["ids_f32"], to=onnx.TensorProto.FLOAT)
+    cast_mask = onnx.helper.make_node("Cast", ["attention_mask"], ["mask_f32"], to=onnx.TensorProto.FLOAT)
+    mul = onnx.helper.make_node("Mul", ["ids_f32", "mask_f32"], ["masked_f32"])
+    reduce_sum = onnx.helper.make_node("ReduceSum", ["masked_f32"], ["sum_f32"], axes=[-1], keepdims=0)
+    cast_out = onnx.helper.make_node("Cast", ["sum_f32"], ["scores_int"], to=onnx.TensorProto.INT64)
+    squeeze = onnx.helper.make_node("Squeeze", ["scores_int"], ["scores"], axes=[-1])
+
+    graph = onnx.helper.make_graph(
+        [cast_ids, cast_mask, mul, reduce_sum, cast_out, squeeze],
+        "causal-masked-equivalence",
+        inputs,
+        outputs,
+    )
+    model = onnx.helper.make_model_gen_version(graph, opset_imports=[onnx.helper.make_opsetid("", 19)])
+    onnx.checker.check_model(model)
+    graph_path = tmp_path / "equiv.onnx"
+    graph_path.write_bytes(model.SerializeToString())
 
-    values = [_ValueInfo("ids", [1, "options", "sequence_length"]), _ValueInfo("mask", [4, 32, 512])]
-    infos = _value_infos(values, _FakeOnnx())
+    session_options = ort.SessionOptions()
+    session_options.log_severity_level = 3
+    providers = ["CPUExecutionProvider"]
+    session = ort.InferenceSession(str(graph_path), session_options, providers=providers)
 
+    live_tokens = [101, 2023, 2003, 102]
+    n_live = len(live_tokens)
+    n_options = 2
+    for width in (64, 128):
+        assert n_live <= width
+        ids_flat = live_tokens + [0] * (width - n_live)
+        mask_flat = [1] * n_live + [0] * (width - n_live)
+        ids = [ids_flat[:], ids_flat[:]]
+        mask = [mask_flat[:], mask_flat[:]]
+        import numpy as np
+
+        feed = {
+            "input_ids": np.array([ids], dtype=np.int64),
+            "attention_mask": np.array([mask], dtype=np.int64),
+        }
+        (scores_out,) = session.run(["scores"], feed)
+        assert np.all(np.isfinite(scores_out)), f"width={width} produced non-finite output"
+        assert scores_out.shape == (1, n_options), f"width={width} unexpected shape {scores_out.shape}"
+        expected_sum = sum(live_tokens)
+        assert np.all(scores_out == expected_sum), f"width={width} scores={scores_out} expected={expected_sum}"
+
+    infos = _value_infos(session.get_inputs(), onnx)
     assert infos[0]["shape"] == [1, "options", "sequence_length"]
-    assert infos[1]["shape"] == [4, 32, 512]
-    assert infos[0]["name"] == "ids"
+    assert infos[1]["shape"] == [1, "options", "sequence_length"]
 
 
 def test_variable_export_sequence_width_64_and_128(tmp_path: Path) -> None: