indexing an array with decreasing time returns array with time dimension even for a single time selected

I believe I have found an additional bug related to non-increasing time (whether it needs to be monotonic I do not know).
I didn't know how to formulate the title in a more concise way, apologies.

Background: My own research data has decreasing time as standard output (back trajectories), but I will present this example using xr's tutorial data. I already found the bug that you can't select a forward going time slice in a backwards going data set. (I can't find the bug report anymore, sorry).

I know the work-around is easy, and it can be discussed if slices with labelled dimensions should support slices in either direction regardless of the direction of the dataarray too.
But that a single selection still returns a time dimension surely is unintended behavior.

What happened:
When selecting a single time step in a decreasing DataArray, the returned DataArray still has a time dimension, of size one.
This makes trouble for different situations that expects 2D data: for instance plot.contourf() which does not reduce the time dimension - as apparently .plot() does.

What you expected to happen:
A returned DataArray with no time dimension.

Minimal Complete Verifiable Example:

ds = xr.tutorial.open_dataset("air_temperature.nc")
print(ds.air.sel(time='2014-12-31T18').shape)
>> (25, 53)
ds=ds.reindex(time=list(reversed(ds.time.values))) #reverse time to show effect
print(ds.air.sel(time='2014-12-31T18').shape)
>> (1, 25, 53)

Environment:

Reproduce the behavior with DataArray.sel after reversing the time coordinate using reindex, starting from the selection path involved in .sel. Add a regression test showing that selecting one timestamp from decreasing time returns no time dimension, and verify the tutorial-data example and plot.contourf use case.