Extract row normalization policy

docs/prompt-architecture-improvement-plan.md

@@ -165,6 +165,10 @@ Already isolated:
   rows now emit `action_family`, `position_family`, `position_key`, and
   `position_keys` so formatter routing and debugging do less keyword guessing.
   Krea, SDXL, and training-caption routes consume these fields when present.
+- final row and pair text normalization lives in `row_normalization.py`,
+  covering trigger prepending, extra-positive append, negative merge/dedupe,
+  caption-part joining, and embedded soft/hard row sanitation before metadata
+  leaves generation.
 
 ### Pair / Adapter Layer
 
@@ -198,7 +202,7 @@ Already isolated:
   root clothing-state assembly.
 - final pair output assembly lives in `pair_output.py`, including soft/hard
   prompt strings, trigger preservation, negatives, captions, and root metadata
-  shape.
+  shape; the final cleanup step is delegated to `row_normalization.py`.
 
 ### Krea2 Formatter Path