I would add to this list, if we haven't done it already:

• Avoid using PyVista's toolbar and menu bar and create/use our own as needed

Done in here on this PR:

Update after dd0d67d (with experimental margin=False):

I see the main challenges being 1) ensuring correct size and 2) managing picking nicely for all the viewers

Can you check to make sure volumetric rendering looks okay? Just want to make sure it plays nicely with the ambient occlusion etc.

There are some issues with the colorbar:

BackgroundPlotter has no attribute named scalar_bar

SSAO=False	SSAO=True

import os.path as op
import mne
data_path = mne.datasets.sample.data_path()
sample_dir = op.join(data_path, 'MEG', 'sample')
subjects_dir = op.join(data_path, 'subjects')
inv = mne.minimum_norm.read_inverse_operator(op.join(
    sample_dir, 'sample_audvis-meg-vol-7-meg-inv.fif'))
    # sample_dir, 'sample_audvis-meg-oct-6-meg-inv.fif'))
evoked = mne.read_evokeds(op.join(sample_dir, 'sample_audvis-ave.fif'))[0]
evoked.apply_baseline((None, 0)).pick_types(meg=True)
stc, res = mne.minimum_norm.apply_inverse(
    evoked, inv, return_residual=True, pick_ori='vector')

initial_time = 0.1
# stc, kwargs = stc, dict(overlay_alpha=0.5, brain_alpha=0.5, vector_alpha=0.)
stc, kwargs = stc.magnitude(), dict(alpha=0.5, surface='white')
kwargs['src'] = inv['src']
meth = stc.plot_3d if isinstance(stc, mne.VolSourceEstimate) else stc.plot
brain = meth(
    subjects_dir=subjects_dir, initial_time=initial_time,
    clim='auto', hemi='both', # views=['sag', 'axi', 'cor'],
    view_layout='horizontal', smoothing_steps='nearest', verbose=True,
    time_viewer=True, size=(800, 600), background='k',
    show_traces=False,
    **kwargs)

and plot_beamformer_lcmv:

It's better after #9121

This PR now requires pyvista/pyvistaqt#90.

I plan to move the mplcanvas to a bottom dock and use the new resize_content() function to replace ensure_minimum_size.

In my experiments, the mplcanvas as a dock has several advantages:

• it can be made separate dynamically with the default dock features (meaning that the "separate" particular case can be removed)
• it does not mess with the central widget layout meaning that the resize_content() function is enough to ensure that the viewer has correct dimensions (important for screenshot). I have to demonstrate that by updating the tests.
• the size of the dock can be handled separately

This is how it looks locally:

Probably the margins can be nullified.

Agreed docking is useful, can that be a first PR, separate from the MultiPlotter stuff? It seems orthogonal, easier to review, and will get in sooner than if we tie it into this PR

Circle does not seem ready yet. I'll investigate locally:

RuntimeWarning: No 3D backend found. Resorting to matplotlib 3d

Closing for now. It already served the purpose of refactor _Renderer and the pyvista 3d backend. I can reopen when I get back to SSAO.