fix(s31): close 2 review findings — drop description from the embedding + re-derive SC2 on the minted shape

tests/test_semretrieval.py

@@ -77,17 +77,36 @@ def test_fake_embedder_is_insensitive_to_code_set_ordering() -> None:
 
 
 def test_distinct_features_give_distinct_vectors() -> None:
+    """SC4, NARROWED: "distinct" means distinct in the EMBEDDED fields — the structural triple.
+    The description-only variant moved to the test below, which asserts the opposite."""
     embedder = FakeEmbedder()
     baseline = embedder(_features())
     for variant in (
         _features(codes=frozenset({"09.1"})),
         _features(measure_type="rate_renegotiation"),
         _features(saving=900_000.0),  # different magnitude bucket
-        _features(description="totally different wording"),
     ):
         assert not np.array_equal(baseline, embedder(variant))
 
 
+def test_features_differing_only_in_description_embed_identically() -> None:
+    """SC4's narrowing, asserted POSITIVELY rather than left implicit — this is the same
+    assertion the loop above used to carry, with the opposite sign.
+
+    ``description`` is excluded from the embedding on purpose: ``verdicts.similarity`` already
+    ignores text by design and ``verdicts._mint_id`` already hashes only the structural triple, so
+    two proposals differing only in prose are ONE verdict as far as the rest of the system is
+    concerned. Leaving the embedding as the sole place surface text still counted made the
+    framework's own restatement of a query outrank real expert prose (defect P1).
+
+    Detach point: put ``features.description`` back into ``_canonical_feature_string`` → RED."""
+    embedder = FakeEmbedder()
+    assert np.array_equal(
+        embedder(_features()),
+        embedder(_features(description="totally different wording")),
+    )
+
+
 def test_fake_embedder_components_are_non_negative() -> None:
     """Non-negative components keep ``cosine`` in [0, 1], matching the structural score's
     range — no scale mismatch when the two are blended by HybridRanker."""