I recently stumbled on this tool.
selimanac.github.io/physics-body-editor
And it also has an option to "guess polygon shape" based on an input image. Except the result... it's just infinitely better.
There's even some sliders to adjust sensitivity and alpha threshhold. I know Construct has kind of a soft limit of 16 points and it keeps (annoyingly) yelling at me when I add more but sometimes you just need more. Is that limit still relevant in modern times and supposing I don't do collision checks on thousands of instances?
Either way, it's clearly possible to improve the results massively.
EDIT: Sometimes I also do get really weirdly bad results.