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

src/portfolio_optimiser/semretrieval.py

@@ -80,14 +80,26 @@ def _magnitude_bucket(value: float) -> int:
 
 
 def _canonical_feature_string(features: ProposalFeatures) -> str:
-    """Order-independent textual projection of the features. ``affected_codes`` is a set, so it
-    is sorted; the raw saving is bucketed so near-identical amounts land on the same point."""
+    """Order-independent projection of the STRUCTURAL triple: sorted ``affected_codes``, the
+    measure type, and the magnitude bucket. ``affected_codes`` is a set, so it is sorted; the raw
+    saving is bucketed so near-identical amounts land on the same point.
+
+    ``description`` is deliberately EXCLUDED, matching the two invariants that already govern this
+    data: ``verdicts.similarity`` ("text is ignored by design") and ``verdicts._mint_id``, which
+    hashes the same three fields — so two proposals differing only in prose already share one id
+    and one structural score. Including description here made the embedding the single place where
+    surface text leaked into retrieval, which inverted the feature's purpose: on a mixed store the
+    framework's own echo of the query outranked genuine expert prose.
+
+    Honesty limit, stated at the site: over a structural tie the cosine ordering is deterministic
+    but semantically ARBITRARY, because the shipped ``FakeEmbedder`` is a sha256 projection with no
+    semantics. Retrieval *quality* arrives only with an injected embedder — the seam is the
+    deliverable, not better ranking."""
     return "|".join(
         (
             ",".join(sorted(features.affected_codes)),
             features.measure_type,
             str(_magnitude_bucket(features.claimed_saving_nok)),
-            features.description,
         )
     )