Cure for feature flags inconsistency
It happens too often that the assertions about Project properties that depend on the feature flags that the test preparation just set up fail, and this should be enough to prevent it. It seems to be simply due to per-process RAM caches ("memoization").
The price to pay (1 minute of waiting) is acceptable in CI (compared to a full retry). In development, it should be bearable, because a dev session typically doesn't change stuff like --heptapod-hg-native
values at each run.
In the course of investigation, lots of improvements on old methods not leveraging utilities such as wait_assert()
.