Oh, Jim Jewett found a bug: if set_traceback_limit() is called while tracemalloc is tracing, two tracebacks are seen different because their length is different, whereas the full traceback would be the same.

To avoid this issue, I propose to add an optional nframe parameter to the set_traceback_limit() function.

With this API, it may be possible to remove the arbitrary limitation of 100 frames by allocating a buffer in start().