It isn't 100% obvious that the new repr for Categoricals is an improvement, but it's non-crazy. One of the remaining xfails one is for eval(repr(categorical_index)) round-tripping that won't be fixable unless we revert back to the old repr behavior.

I'm pretty sure that the fix in test_astype_dt64_to_string is correct and the test is just wrong, but merits a close look.

That leaves 12 xfails, including the one un-fixable round-trip one that we'll just remove. Of those...

• test_join_on_key i think is surfacing an unrelated bug that I'll take a look at
• test_to_dict_of_blocks_item_cache is failing because we don't make series.values read-only for ArrowStringArray. I think @mroeschke can comment on how viable/important that is.
• test_string_categorical_index_repr is about CategoricalIndex repr that span multiple lines; with the StringDtype the padding is changed.
• 4 in pandas/tests/io/json/test_pandas.py that im hoping @WillAyd can take point on
• test_to_string_index_with_nan theres a MultiIndex level that reprs with a nan instead of NaN. Not a huge deal but having mixed-and-matched nans/NaNs in the repr is weird.
• test_from_records_sequencelike i don't have a good read on
• tests.base.test_misc::test_memory_usage is skipped instead of xfailed, but the reason says that it "doesn't work properly" for arrow strings which seems xfail-adjacent. Instead of skipping can we update the expected behavior cc @jorisvandenbossche ?

(Update: looks like I missed one in test_http_headers and another in test_fsspec)