fix(segmentation): hash the extracted text and let the plan key fire

src/llm_ingestion_okf/inbox.py

@@ -41,6 +41,7 @@ from .segmentation import (
     SegmentationPlan,
     SegmentEntry,
     assert_plan_applies,
+    observed_extractor_version,
     slice_segments,
 )
 from .structure import (
@@ -411,17 +412,22 @@ def _render_segments(
     A refusal is therefore reported once, for the document, rather than once
     per segment: the operator's unit of review is the document they dropped.
     """
+    extractor_id = Path(path.name).suffix.lower().lstrip(".") or "none"
     assert_plan_applies(
         plan,
         source_sha256=hashlib.sha256(source_bytes).hexdigest(),
-        extractor_id=Path(path.name).suffix.lower().lstrip(".") or "none",
-        # The plan's own value, passed through. Door B can observe WHICH
-        # extractor ran (the suffix is what dispatches it at `extract.py`) but
-        # not the version of a third-party parser -- `pdfplumber`'s transitive
-        # `pdfminer.six` pin is the measured example. Naming the key and
-        # leaving its value to whoever knows it is the same division D5 makes
-        # for `bundle_id`; a fabricated value here would make S5b decorative.
-        extractor_version=plan.extractor_version,
+        # `text` is the canonical extracted text this run produced, AFTER the
+        # profile's renderer -- exactly the string the plan's offsets index.
+        text_sha256=hashlib.sha256(text.encode("utf-8")).hexdigest(),
+        extractor_id=extractor_id,
+        # OBSERVED, never the plan's own value passed back in. That is what
+        # this line used to do, and comparing a value with itself made the
+        # version half of S5b unable to fail: a plan adjudicated under one
+        # converter replayed silently under another. The version of a
+        # third-party parser is knowable here after all -- `pdfplumber`'s
+        # transitive `pdfminer.six` pin is the measured example, and
+        # `observed_extractor_version` is where each row names its source.
+        extractor_version=observed_extractor_version(extractor_id),
     )
     sliced = slice_segments(text, plan)