Ran into almost the exact same wall with an intake classifier last year. The thing that finally broke the cycle for me was treating "Other" as a signal instead of a category — when confidence was low I routed it through a second pass that had the model summarize the email in one line and re-classify with that summary attached. Cut the dumping-ground effect dramatically. Also found that fewer, longer examples beat longer definitions; the model learned the category boundaries better from edge cases than from prose. Solid write-up — the debugging metaphor is exactly right.