`on(sch,sndr)` and `starts_on(sch,sndr)` should not transform the senders they returns

[ericniebler]

the overload of on that takes a scheduler and a sender (on(sch, sndr)) is defined to be:

transform_sender(
  query-or-default(get_domain, sch, default_domain()),
  make-sender(on, sch, sndr))

so we use the domain of the scheduler to find a customization. that's odd, because that is neither where the on sender will start nor where it will complete. on(sch,sndr) does not know where it will be started, and since it completes where it starts, it doesn't know where it completes either.

other algorithms use the domain of the predecessor sender. on(sch,sndr) has no predecessor sender. the logical conclusion is that on should not transform the result sender.

on(sch,sndr) should be simply make-sender(on, sch, sndr).

the same reasoning also applies to starts_on.

[ericniebler]

OTOH, starting work on an execution context is a special case, and it should be possible to customize [starts_]on for a particular domain.

closing nad