Housekeeping left over from the #293 vocabulary rename #356

Minor findings from PR #354's review, deferred so that PR stayed reviewable. None affects behaviour.

Bridge / API surface

• nameparser/config/particles.py and bound_given_names.py have no __all__, so from ... import * leaks assert_normalized, and particles.py re-exports BOUND_GIVEN_NAMES (which belongs to the other module and is imported only for its disjointness assert). Their 1.x predecessors gained __all__ in Rename the vocabulary data modules to the 2.0 terminology, with a 2.x bridge (#293) #354; the migration destinations did not.
• __dir__ raises KeyError rather than AttributeError if the module has been dropped from sys.modules. Off the read path — only an explicit dir() reaches it.
• inspect.getmembers() / help() on titles/suffixes resolves the retired names and emits their DeprecationWarnings.

Tests

• _RETIRED_NAMES in tests/v2/test_config_aliases.py is fully derivable from ALIASES (verified identical). A second hand-written copy of the same five names in the same file is drift surface: a sixth alias added to one and not the other silently exempts that name from the internal-usage scan.
• ALIASES is a bare 4-tuple indexed positionally in six places (for m, n, _, _ in ALIASES). _LatinCopy in tests/v2/test_ledger_guards.py is the suite's own NamedTuple precedent.
• test_star_import_binds_exactly_the_live_and_retired_names asserts names but never values, so an __all__ entry routed to the wrong constant is caught only incidentally.
• from X import Y warning attribution is tested for one alias of five. The two same-module aliases are a structurally different shape (the table sits at the bottom of a module that is mid-execution during import).

Docs

• docs/customize.rst says "TITLES, PARTICLES and the rest of nameparser.config" are frozen — CAPITALIZATION_EXCEPTIONS and REGEXES are still dicts — and its next sentence reads as though TITLES was renamed, which it wasn't.
• AGENTS.md:208 keeps "first-name join"/"last-name join" on a line whose constants were renamed to the given/family vocabulary.
• AGENTS.md:240 says SUFFIX_ACRONYMS − {esq} "changes many"; measured, only the multi-dot spellings change. The paragraph's refusal to publish a count is right — "changes the multi-dot spellings" is both truer and count-free.
• AGENTS.md:254's nickname_delimiters string-sentinel gotcha is stale for 2.x: CONSTANTS.regexes.parenthesis = ... raises TypeError, and delimiters reach the parse as Policy pairs. It sits in a section not marked as v1 history.
• nameparser/config/__init__.py cites the migration spec's §3 for a list that lives in §4, and both citations point into gitignored docs/superpowers/, so nobody without that directory can check either.

Found by the five-agent review on PR #354 (comment accuracy, test coverage, silent failures, type design, general code review).

Added after the 2.2 bridge changed semantics

• AGENTS.md:224 still explains the internal-usage guard by saying an internal read "would also consume the once-per-process warning and leave the real caller told nothing." That stopped being true when the bridge moved to per-read-location warnings. Measured on master:

  internal_reader.py:1  then  real_caller.py:1  ->  2 warnings
  

  The guard is still worth keeping — an internal read emits a warning attributed to library code the caller cannot act on, and breaks their -W error::DeprecationWarning runs — but the stated reason is wrong. docs/release_log.rst:48 states the new semantics correctly, so this is the only stale site.