Answer architecture round 2: one folder walk and one sorted-keys helper, REGENERATE for the index, and the held decisions naming the items that end them

README.md

@@ -816,8 +816,8 @@ The target is about a microsecond per value (goal 12).
 Each file is parsed, validated and weight-indexed once, bar a table's name and parent
 columns, which are mapped on the first `Fake` that selects by name or descends through
 it. With the shipped set alone, a category loads on the first call reaching it, with
-every category it reads, and `List` reads an index; beside a `--data-path`, every
-category loads in `New`. Proving the draw fences adds a pass over the loaded tree per
+every category it reads and every table linked to it, and `List` reads an index for what
+has not loaded yet; beside a `--data-path`, every category loads in `New`. Proving the draw fences adds a pass over the loaded tree per
 fence, and walks what a render reads only where data holds a reference, so a set that
 holds none pays for the passes alone. A `Fake` call then costs about what its output
 costs: an unweighted pick is O(1) whatever the list's length, a weighted one