test_large_model_generation_is_not_quadratic trips its 20 s budget on CI runners (22.7 s, 22.2 s) #1479

python/tests/test_codegen_gen_scaling.py::test_large_model_generation_is_not_quadratic times generate_combined_from_model on a 40,000-reaction model against a 20 s wall-clock budget. Its docstring says the budget has ">20x headroom on a normal machine, so it trips only on a genuine re-quadratic regression, not on CI jitter". That is true locally (1.7 s on an M-series Mac), but not on GitHub's ubuntu runners:

• 22.7 s: run 40425653155, branch fix/1298-sens-rhs-clamp-retry, 2026-10-27
• 22.2 s: run 41647495310, PR bngl: warn when a source's first run starts from a state from_bngl drops (#1366) #1388, 2026-11-03 (the "Python suite · ubuntu-latest · 1/2" shard)

Neither change touched codegen. Both runs went green on gh run rerun --failed. The shard is a required check, so each trip blocks a merge until someone reruns it.

On those runners the headroom is about 1x, not 20x, so the test is a coin flip on a slow machine.

Fix direction: measure scaling, not seconds. Time the same generator on a model a tenth the size, in the same process, and bound the ratio: linear is about 10, quadratic about 100, so a bound near 30 separates them on any runner. The alternative is to raise the budget, which only moves the coin flip.