Tell a system to totally redo code that sits in its context and it hands back a
paraphrase. It relabels a name, swaps two lines, and claims a redo.

Command force does not repair this. A prompt that names the failure ahead of time
and forbids cosmetic tweaks by name still gets cosmetic tweaks. Three drivers push
it, and none of them answer to phrasing.

1. The original is a lodestone. Output leans on the text in view.
   A veto does not outweigh that pull. Deletion does.
2. Volume shrinks care. A redo costs far more output than a nudge. Given
   forty items in one pass, the system spreads its energy and every item gets a nudge.
3. Negative rules offer nothing to aim at. "Not a narrator" and "no longer A*"
   name what to dodge. With no positive mark, the closest concrete artifact is
   the old code, so the output stays near it.

This method strips the text, breaks the work into single dispatches, and states
the goal in positive terms. It then weighs the result against the deleted
original, because a system that faked this way once will claim victory again.