fix(corpus): harden canonical chunk and frontmatter invariants

harness/tests/test_corpus_normalize.py

@@ -52,6 +52,29 @@ def test_frontmatter_can_end_at_eof_without_inventing_content():
     assert document.content == ""
 
 
+@pytest.mark.parametrize(
+    "frontmatter",
+    [
+        "title: first\ntitle: second",
+        "title: &shared value\ncopy: *shared",
+        "nested: " + "[" * 25 + "x" + "]" * 25,
+        "items: [" + ",".join("x" for _ in range(1100)) + "]",
+        "api_key: secret",
+    ],
+)
+def test_rejects_unsafe_frontmatter_as_typed_permanent_error(frontmatter):
+    raw = f"---\n{frontmatter}\n---\nbody".encode()
+    with pytest.raises(PermanentNormalizationError) as caught:
+        normalize(acquired(raw), "pipe:v1")
+    assert caught.value.reason == "invalid_frontmatter"
+
+
+def test_pipeline_policy_errors_are_not_misclassified_as_bad_frontmatter():
+    with pytest.raises(ValueError, match="pipeline_version") as caught:
+        normalize(acquired(b"---\ntitle: valid\n---\nbody"), "")
+    assert not isinstance(caught.value, PermanentNormalizationError)
+
+
 @pytest.mark.parametrize(
     ("content", "media_type", "reason"),
     [