@network-harness/executor-replay

The executor seam's test double: scripted replies, in order, with no model.

Everything the loop is responsible for -- the output contract, the one correction turn, budgets, the terminal event -- has to be assertable without a model, or it would only ever be tested by watching a real run and hoping. A script of one entry per turn makes "the model replied badly, then well" an ordinary test.

It also keeps every prompt it was handed, which is how a test asserts what the model was actually told.