The work-requesting schedulers steal across NUMA domains regardless of --hpx:numa-sensitive / hpx.numa_sensitive.

• local_workrequesting_scheduler::set_scheduler_mode (local_workrequesting_scheduler.hpp:1993) always adds enable_stealing_numa. The thread manager removes it again with update_scheduler_mode for a nonzero level, but the scheduler never reads the flag.
• Each worker's victim mask only contains the worker itself (:1943), and next_victim (:1635) picks a random worker out of all of them. There is no preference for workers in the same NUMA domain and no way to restrict stealing to it.

For comparison, the local-priority scheduler builds a victim list per worker in on_start_thread (local_priority_queue_scheduler.hpp:1649): workers in the same NUMA domain first, then (if enabled) workers in other domains. #7756 restores the three sensitivity levels (0 - no restriction, 1 - only boundary workers steal from the nearest domains, 2 - no stealing across domains) for the local and local-priority schedulers.

The work-requesting schedulers should at least prefer victims in the same NUMA domain and honor level 2 (no stealing across domains). Until then the documentation of --hpx:numa-sensitive should say that the setting has no effect for them.