Aside: bisecting flakes doesn't have to involve repeated runs. You can reformulate bisection as an information probing operation, expanding the scope to support noisy benchmarks or low-probability flakes. Bayesian inference narrows down the probable range of the failure for each new observation, and you can choose new probes to maximize information gain-- or even run them in parallel to minimize the total time.

You do have to provide flake rate probability to do the probability estimates, but even roughly correct rates work fine. Running bisects assuming a 5% chance of getting a false positive or negative barely adds more steps and greatly improves robustness.

The math is pretty simple too-- my old prototype might still be in Google's monorepo; I should reimplement it for the open source world.