The hierarchy in §V is doing a lot of work for an empirical claim.
"Mechanisms override reasoning, which overrides character." Any case where a person dominates can be filed as short horizon or as a mechanism not yet identified. Any case where architecture dominates counts as confirmation. That is a comfortable likelihood function.
The inspector, teacher, and politician show that incentives can swamp intention. They do not show that this is the generic order of causal strength, or how you would know if it were not.
The page already tries to bound this. Character can dominate an episode; the claim is about distributions over succession. It also says the derivation chain breaks if selection is not the primary filter.
So the identification is supposed to be the horizon: longer than tenure, architecture is load-bearing; shorter, a person can be decisive.
Horizon-talk is a classification, not a test.
You still need a comparison that could go the other way: hold formal incentives, feedback, and selection rules approximately fixed and vary occupant character; then hold the occupant pool approximately fixed and vary architecture. If the first move keeps shifting the distribution and the second does not, mechanism-primacy should be downgraded rather than repaired by finding a subtler mechanism.
Until that comparison is specified, "empirical claim about causal primacy" is a label.