This PR is my proposal for fixing #1904 (previous attempt 431ec6d was reverted).

Root cause analysis

The main entry for rendering the traceback for a test exception is Node._repr_failure_py. It takes the excinfo and returns a TerminalRepr. Two parts of it are relevant here:

This calls the node's _prunetraceback method with the excinfo, which then proceeds to filter out internal entries, hide __tracebackhide__ = True frames, etc. It modifies the excinfo in-place by updating its excinfo.traceback field.

This calls into the main FormattedExcinfo machinery to do the bulk of the work, but passes tbfilter=False because it already did the filtering itself, in a Node-customizable manner, so doesn't want the generic filtering that FormattedExcinfo does.

Next, need to look at how FormattedExcinfo handles exception chaining.

It starts with the excinfo it was passed, which is the main exception. It renders it, then checks if it has any chained exceptions (either __cause__ or __context__), and if so, creates an ExceptionInfo for them, renders them, and does the same in a loop.

If you've followed closely you see the issue - the main exception got the filtering treatment from the Node's _prune_traceback before it was even passed to FormattedExcinfo. But the chained ExceptionInfo's are created inside FormattedExcinfo, and we've passed tbfilter=False so they don't even get the basic filtering.

Proposed solution

I think it would have been best if the chained ExceptionInfo were represented in the ExceptionInfo itself, like how base exceptions work. Then the _prunetraceback stuff would also prune them before passing to FormattedExcinfo and all is well. But this is somewhat difficult to achieve, so I went with something easier but less clean.

Instead of doing the _prunetraceback before FormattedExcinfo, I allow tbfilter to also be a callback in addition to the existing False/True. Then we pass tbfilter=self._prunetraceback instead of calling it ourselves and passing tbfilter=False. This then causes the chained exceptions to be filtered exactly like the main exception.

Technical notes

As mentioned above, the _prunetraceback mutates the excinfo in-place. I really didn't like this behavior for the tbfilter callback. So instead, I changed it to be (ExceptionInfo) -> Traceback. This breaks the _prunetraceback interface, so I renamed it to _traceback_filter.

There is also some internal refactoring to make this possible, namely removing the ghastly WeakReference cycle which makes TracebackEntry know about its ExceptionInfo, and also making TracebackEntry immutable. These are separate commits.