image search test

docs/SCAN_TO_ORDER.md

@@ -244,7 +244,23 @@ you change ranking — it is the spec.
 
 Scores: vector = `1 − cosine distance`; text = 0.95 for the whole label
 inside the name, else `0.8 × (label words found / label words)`; combined =
-`max(vector, text) + 0.10` when both hit, capped at 1.
+`max(vector, text) + 0.10` when both hit, capped at 1. Ties are broken by
+cosine distance — nearest first, a text-only row last — and only then by
+name.
+
+The label and the product name are both separator-folded before that
+substring test (`utils.FoldSeparators`), and compared again with separators
+removed (`utils.TightenLabel`, labels of 4+ characters), so the brand's own
+punctuation does not decide the match: "Parle G", "Parle-G" and "ParleG" all
+reach *Parle-G Original Glucose Biscuits*. A single-character token survives
+tokenising when it follows a word, because it is often the whole name — the
+"G" of Parle-G, the "K" of Special K. It is still dropped when it stands
+alone or is a pack multiplier.
+
+All three mattered at once: before this, "Parle G" tied with *Parle Monaco
+Classic* at 0.9 (the "G" was dropped, so only "parle" matched either row),
+and the name tie-break handed it to Monaco because a space precedes a hyphen
+in ASCII. A confident, wrong answer — the kind no score floor can catch.
 
 ### Changing the embedding model