Enrich formatter route trace metadata

docs/prompt-architecture-improvement-plan.md

@@ -29,6 +29,10 @@ The map audit currently sees:
 - Pair seed simulation, so Insta/OF soft/hard metadata and formatter outputs
   prove locked determinism and person/scene/content/pose/expression/
   composition reroll behavior.
+- Formatter route traces expose selected metadata fields, so Krea2, SDXL, and
+  caption outputs can be debugged by category, action/position family, selected
+  pair side, scene profile, position keys, and POV labels instead of only
+  proving that a metadata branch was used.
 
 ## Architectural Finding
 

docs/prompt-pool-routing-map.md

@@ -242,7 +242,10 @@ and to compare the exact per-axis RNG seed used for the row.
 `tools/prompt_map_audit.py` includes a runtime metadata route check. It builds a
 representative single row and Insta/OF pair, verifies embedded
 `generation_trace` fields, and confirms Krea2, SDXL, and caption formatters
-consume metadata JSON instead of silently falling back to raw prompt text.
+consume metadata JSON instead of silently falling back to raw prompt text. The
+formatter route traces also expose selected row metadata such as selected pair
+side, category, action/position family, scene profile, position keys, and POV
+labels.
 The same audit also statically rejects direct `row["prompt"]` reads in
 formatter metadata modules outside the shared fallback helpers.
 
@@ -997,6 +1000,9 @@ them through Krea2, SDXL, and training-caption paths, and reports structured
 issues for:
 
 - formatter routes falling back away from metadata;
+- formatter route traces exposing selected row metadata such as category,
+  action/position family, selected pair side, scene profile, position keys, and
+  POV labels;
 - raw builder labels leaking into Krea output;
 - duplicate negative-prompt comma items;
 - softcore prompt noise;