fix(s31): close 1 review MAJOR — bound the blend weight so the hybrid preserves structural order

src/portfolio_optimiser/semretrieval.py

@@ -78,8 +78,29 @@ if TYPE_CHECKING:  # verdicts imports agent_framework — keep it out of the run
 EMBED_DIM = 64
 
 # Blend weight for HybridRanker: score = weight * cosine + (1 - weight) * structural.
-# 0.5 gives the two falsifiable signals equal say; the caller can override per construction.
+# The caller can override per construction.
-SEMANTIC_WEIGHT_DEFAULT = 0.5
+#
+# 0.25, and the reason is MEASURED, not aesthetic. The previous value (0.5) was justified as
+# "giving the two falsifiable signals equal say" — a claim about equivalence that the measurement
+# refutes. Over the 1770 pairs of a fully enumerated 60-feature family, the ``FakeEmbedder``
+# cosine is a near-constant: mean 0.7500, sd 0.0380, range [0.6123, 0.8614] => spread 0.2492. A
+# term with that little dynamic range is a TIE-BREAKER, not a co-equal signal, and weighting it as
+# though it were an equal partner does not make it one — it just lets 0.25 of a near-constant
+# outvote real structural distance.
+#
+# Worst-case bound: the semantic term can overturn a structural gap G only when
+# w * spread > (1 - w) * G, i.e. it is safe while ``w < G / (G + spread)``. With G = one
+# ``verdicts._W_MAGNITUDE`` step (0.15) and the measured spread, that ceiling is 0.3758 — so 0.5
+# sat ABOVE it and the ordering was genuinely overturnable; 0.25 sits below with margin. Stated as
+# a property rather than a fixture fit: w = 0.25 holds for any spread below 0.45, against 0.2492
+# observed. Measured, not assumed: 9 adverse triples at w = 0.5, zero at every weight from 0.45
+# down to 0.05 (see ``test_hybrid_preserves_structural_order_across_one_magnitude_step``).
+#
+# Lowering it costs nothing the seam is for: cosine still decides 100% of exact structural ties at
+# any w > 0, which is the role SC2 pins. Only the co-equal-signal role disappears, and it was
+# never defended. An injected REAL embedder has a different cosine distribution and therefore
+# needs its own weight justification — this number is derived from the shipped fake one.
+SEMANTIC_WEIGHT_DEFAULT = 0.25
 
 # Mirrors ``verdicts._MAGNITUDE_BUCKETS``. DUPLICATED, not imported, on purpose: importing it
 # would pull ``verdicts`` (and therefore ``agent_framework``) into this module's runtime graph
@@ -121,7 +142,13 @@ def _canonical_feature_string(features: ProposalFeatures) -> str:
     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."""
+    deliverable, not better ranking.
+
+    And the limit is wider than "ties": an arbitrary term does not politely confine itself to
+    exact ties. It also decides NEAR-ties, and ``SEMANTIC_WEIGHT_DEFAULT`` is the only thing that
+    bounds how near — at weight w it can overturn any structural gap below ``w/(1-w)`` times the
+    cosine spread. That is why the weight is derived from a measurement rather than chosen for
+    symmetry, and why raising it is a correctness change, not a taste one."""
     return "|".join(
         (
             ",".join(sorted(features.affected_codes)),