Same machine, one variable, two futures

Two plausible fixes for the memory leak. On most platforms you try one, watch it for a day, then try the other: serially, against a baseline that shifted in between.

The agent forks instead. Two POSTs to /copy, both from the same snapshot, so both containers begin as byte-identical machines. Fix A lands in one, fix B in the other. The agent runs the same test load against both URLs and reads both health endpoints and both memory curves.

This is a controlled experiment in the lab sense: one variable, identical everything else. The copies are copy-on-write clones on the box you already rent, so two futures cost roughly the disk they diverge by.

The loser is a DELETE. The winner is already running. Same machine, one variable, two futures.

Same machine, one variable, two futures.