Bound only the Start press in the language test

Merged
CO
4d ago
+17-8

The python and java variants of the pending-judge test set a 2 s page timeout before the preflight, so every click and wait leading up to Start shared it. On a loaded runner the Start click alone stalled past 2 s and the python case failed on #436, which does not touch this path, while the java case passed in the same run. The bound exists to prove a non-C start skips the 5 s judge wait, so it now covers only the press to the token request, at 4 s, and the preflight keeps the page's own timeout. Making every start wait for the judge still fails both variants.