fix: only count user-code frames toward max_stack_depth

[irajaryan]

Closes #378.

record_trampoline_hit previously decremented the max_stack_depth budget for every frame walked, including its own frame, the injected _mutmut_trampoline, and any third-party / stdlib frames. This meant a configured value of N didn't correspond to anything a user could reason about.

This change only counts frames whose file lives under the mutants/ working directory and whose function name is not _mutmut_trampoline. os.path.realpath is used so the comparison works on macOS where /var is a symlink to /private/var.

The e2e config project's max_stack_depth is updated from 8 to 5 to keep the same coverage pattern under the new counting; the snapshot is unchanged.

Six unit tests cover: budget disabled, within/over budget, third-party frames excluded, trampoline frames excluded, and the existing test-runner heuristic preserved.

[Otto-AA]

Thanks for the PR, I didn't see it and worked on this simultaneously together with some other changes in #514 . There I decided to only include stack frames part of the source_paths config, so test setup is also not counted.

os.path.realpath is used so the comparison works on macOS where /var is a symlink to /private/var.

That's something I did not think about in my implementation. Resolving symlinks could indeed be necessary.

[irajaryan]

Thanks for the PR, I didn't see it and worked on this simultaneously together with some other changes in #514 .

Ah, didn't see #514, looks like a much bigger cleanup and I think filtering against source_paths is the right call (it correctly excludes test files, which my approach doesn't).

Happy to close this in favor of #514. Let me know if I need to open a follow-up PR with just the realpath fix once #514 lands, or if you'd prefer to fold it into #514 directly. The relevant bit is small.