The hostile input tests compared elapsed time against fixed limits, so
they measured the machine as well as the code. They fail on slower or
loaded builders, such as the nixpkgs x86_64-linux check phase, even
though rendering stays linear.
Compare the CPU time of an input against one four times as long
instead. Linear rendering gives about 4 and a quadratic regression
about 16. CPU time is needed because a wall-clock ratio inflates under
bursty load, when a short run fits between bursts and a long one does
not.
Rendering past the reference limit gets its own test, since the link
count cannot catch a limit that stops linking but keeps rescanning.
The bare reference test now puts a valid reference before the long
run, so giving up is visible in the result without any timing.
AI-assisted: Claude Code (claude-opus-5-5)
Alessio Caiazza
9c790a7d1c