Comparing clean methods with identical data & parameters (as far as possible): code.

An example using natural weighting:

None of the methods are acceptable. Multiscale is the worst, but hogbom is extremely bad: there are major ringing artifacts (stripes parallel to the main filament).

Disturbingly, even though I have given multiple scales, there are clearly no non-pointlike scales in any of the model images, including mem, so I think there's a major problem with tclean:

I'm re-doing it with uniform weighting, because that's the case I was actually interested in.

The uniform tests come up with a clear and obvious answer: use clark clean, DO NOT use any other methods as they are all unstable.