Two follow-up commits added based on review feedback:

f69ecee — perf: restore _pst cache with proper invalidation

The original fix dropped the cache entirely, which is ~1000x slower per call (~69 µs vs ~70 ns). This commit restores it with correct invalidation:

• SetManager now accepts an optional on_change callback, fired after add / remove — so user mutations immediately clear _pst
• Constants.__setattr__ clears _pst and re-wires the callback whenever one of the four contributing attrs is replaced — this covers the test fixture teardown path where setattr(CONSTANTS, 'prefixes', SetManager(...)) swaps in a fresh instance
• The manual CONSTANTS._pst = None reset in conftest.py is removed; __setattr__ handles it automatically

5fc4b6b — test: expand coverage and add assertNotIn helper

• Added a prime-then-mutate test (_ = c.suffixes_prefixes_titles; c.titles.add(...); assertIn(...)) — the exact scenario from the bug report, which none of the original five tests exercised
• Added add-tests for suffix_acronyms and suffix_not_acronyms (both were untested)
• Added a remove-test for prefixes (mirroring the existing titles remove test)
• Replaced assertFalse(x in ...) with assertNotIn (added shim to HumanNameTestBase)

690 tests pass (682 pre-existing + 8 new × dual-run fixture).

388c65b — refactor: replace the __setattr__ cache-invalidation with a descriptor

A follow-up on the cache-invalidation machinery added in f69ecee. That commit kept the _pst cache correct after add()/remove() by wiring a SetManager.on_change callback and using a custom Constants.__setattr__ (plus a _PST_ATTRS frozenset) to clear _pst and re-wire the callback whenever one of the four contributing attributes was replaced. It worked, but a closer review surfaced three issues worth tidying before merge:

1. The __setattr__ was global for a four-attribute concern. It ran on every attribute assignment to Constants (string_format, capitalize_name, …), paying an in-check each time, and the "which attributes are cached" knowledge was duplicated across three places that had to be kept in sync by hand: the _PST_ATTRS frozenset, the attribute annotations, and the union expression in suffixes_prefixes_titles.

2. A stale-callback leak. Reassigning a manager (c.prefixes = SetManager(...)) wired the new one but never detached the old one. The orphaned manager still held self._invalidate_pst, so mutating that dangling reference would invalidate a cache it was no longer part of — benign (it only nulls the cache, never corrupts it), but a latent foot-gun and an unnecessary recompute.

3. A dead constructor parameter. SetManager.__init__(on_change=...) was never actually passed — wiring always happened afterward via the external value._on_change = ... poke (including the config-teardown setattr path). The advertised constructor contract didn't match how the callback was really wired.

The fix moves the behavior onto a small _CachedUnionMember descriptor placed on exactly the four cached attributes:

• Invalidation/re-wiring now lives beside the attributes it governs and fires only for those four, so the global __setattr__ and the _PST_ATTRS frozenset are both gone.
• On reassignment it detaches the replaced manager (previous._on_change = None), closing the stale-callback leak.
• The unused on_change constructor parameter is dropped; _on_change is initialized to None and wired by the owning Constants, which is the only path that was ever exercised.

The in-place add()/remove() → _on_change callback mechanism on SetManager is unchanged — that's still what catches mutations to an existing manager; only the whole-attribute-replacement path moved from __setattr__ to the descriptor.

Behavior, mypy, and ruff are all unchanged (the descriptor uses @overload for the instance-vs-class __get__ return types so the SetManager type still flows through). Parse throughput is identical to the cached baseline — ~18 µs/name, vs ~263 µs/name with no cache — so the performance reason for keeping _pst still holds.

New tests cover the paths the previous version left unguarded: the replaced-manager re-wiring path, the "orphaned manager no longer invalidates the live cache" guarantee (asserted by cache identity, since the leak's only symptom is an unnecessary recompute), add_with_encoding invalidation, and removal from the two suffix sets after the cache is primed. 700 passed, 20 xfailed.