Motor unit recruitment is dependent upon not only the force required, but also the (intended) speed of movement. That is, the threshold at which a particular motor unit is recruited is lower for “ballistic” than for “ramped” muscle contractions (as Henneman himself termed them in his studies). Applied to cycling, the primary determinant of what fraction of your the pool that is recruited is the power output - whether the cadence (and hence torque) used to generate that power matters is actually still equivocal.
The other factor that matters is time/fatigue. As time passes/the initially recruited motor units fatigue, additional motor units will be recruited to maintain the intensity. Thus, by the end of a prolonged bout of exercise to task failure you will have recruited (and hence trained - as you indicated, adaptations only take place in the muscle fibers that are utilized) all, or almost all, of your motor units (and thus muscle fibers), even if the intensity was actually moderate.
Quadrant analysis is a(n imperfect) graphical tool I came up with about 20 y ago to help visualize such issues. TP seems to have lost the article I wrote describing it, but the same information can be found in TRWPM, or various places on the web.