feat(segmentation): quote anchors so a re-extraction costs a re-anchor

tests/test_propose_segments.py

@@ -142,6 +142,48 @@ def test_the_spans_it_proposes_are_slices_of_the_text_it_read(tmp_path: Path) ->
         assert text[start:end].strip()
 
 
+def test_every_entry_carries_an_anchor_quoting_its_own_span(tmp_path: Path) -> None:
+    """Written at proposal time, when text and offsets are known to agree.
+
+    Reconstructed later it would quote whatever the extraction had already
+    become, which is precisely the drift the anchor exists to survive.
+    """
+    from llm_ingestion_okf.extract import extract_text
+
+    source = write(tmp_path)
+    payload = propose(tmp_path, source=source)
+    text = extract_text(source.name, source.read_bytes())
+    assert payload["entries"]
+    for entry in payload["entries"]:
+        start, end = entry["span"]
+        anchor = entry["anchor"]
+        assert anchor["quote"] == text[start:end]
+        assert text[start - len(anchor["prefix"]) : start] == anchor["prefix"]
+        assert text[end : end + len(anchor["suffix"])] == anchor["suffix"]
+
+
+def test_a_proposed_plan_survives_a_shift_of_the_text_it_was_made_against(
+    tmp_path: Path,
+) -> None:
+    """The round trip the anchor exists for, measured end to end.
+
+    Prepend a paragraph -- exactly the benign edit that leaves every offset
+    short -- and every body must still be the text the proposer quoted. Before
+    anchors this cut each segment early and landed on real prose, with nothing
+    anywhere able to tell.
+    """
+    from llm_ingestion_okf.extract import extract_text
+    from llm_ingestion_okf.segmentation import slice_segments
+
+    source = write(tmp_path)
+    plan = parse_segmentation_plan(propose(tmp_path, source=source))
+    text = extract_text(source.name, source.read_bytes())
+    quoted = [text[start:end] for start, end in (item.span for item in plan.entries)]
+
+    shifted = "Et nytt avsnitt foran alt annet.\n\n" + text
+    assert [body for _, body in slice_segments(shifted, plan)] == quoted
+
+
 def test_the_paths_it_proposes_are_unique_and_hierarchical(tmp_path: Path) -> None:
     plan = parse_segmentation_plan(propose(tmp_path))
     paths = [entry.path for entry in plan.entries]