Finish v2.2.0: trim the release notes, close #433 and #356, sweep the docs

[derek73]

Closes the last two open v2.2 issues, cuts the release notes down to what a caller needs to decide, and runs the release checklist's staleness sweep.

The release notes

The 2.2.0 entry had grown to 12,442 words against 2,222 for 2.1.0 and 2,931 for 2.0.0 — 5.6x the previous minor across fewer bullets, so the bloat was entirely per-bullet (150–771 words each, against 50–150 in 2.1.0). Bullets had become mini design documents: blast radius by baseline, per-field redistribution, gate-blindness disclaimers, recompute recipes, review history.

All of that already lives in docs/design/decisions.md, in more detail — R3 carries #408's parity counts, recompute recipe and gate-blindness note; parse-cost carries the +67 calls and the per-PR attribution; P2 carries #367's grouping argument. So this was deletion plus a pointer, not a migration. Pointers name the entry (the `R3` entry of docs/design/decisions.md), never a line number, following the four bullets that already did.

Each bullet keeps: how a name parsed before, what it does now, the limits a caller would trip over, the 1.4.0 direction, and the issue — which carries the full measurement. 12,442 → 4,927 words.

Closes #433

customize.rst and rules.md's suffix Background both said, truly, that a multi-word vocabulary entry can never match — and neither said the next thing, whose absence is what does the damage: adjacent suffix tokens are reassembled after matching, so a multi-word credential parses as its component words. parse("John Smith, MD PhD").suffix has been 'MD PhD' since 1.4.0.

The inference from "a multi-word ENTRY is inert" to "a multi-word CREDENTIAL is unparseable" survived #291, an approved design spec and a full implementation plan specifying a new vocabulary set, matching unit, predicate and Lexicon field. None of them ran the parse.

Both examples verified non-vacuous: on the default vocabulary "John Smith, LEED AP" gives suffix '' and family 'John Smith'.

Closes #356

Six live items from PR #354's five-agent review, none touching parse behaviour:

• __all__ on the two migration destinations. The 1.x shims got one in Rename the vocabulary data modules to the 2.0 terminology, with a 2.x bridge (#293) #354; particles and bound_given_names did not, so from nameparser.config.particles import * bound assert_normalized and BOUND_GIVEN_NAMES — the latter owned by another module and imported only for a disjointness assert — as though they were vocabulary. This is the one caller-visible change in the PR and it has a release-log bullet.
• __dir__ no longer raises KeyError. It read sys.modules[module], so dir() on a module dropped from sys.modules raised — not among the things dir() may do to a caller. It degrades to the closure's aliases.
• ALIASES is a NamedTuple (for m, n, _, _ in ALIASES appeared six times and reads correctly only if you remember which end is which; _LatinCopy is the precedent), _RETIRED_NAMES is derived from it rather than hand-written a second time in the same file — verified identical to the five rows it replaces.
• The star-import test asserts values, not only names, and the from-import attribution test runs over all five aliases rather than the one prefixes row — the two that kept their module are a different shape, their table sitting at the bottom of a module mid-execution during its own import.

Both fixes mutation-checked rather than merely covered: routing prefixes.PREFIXES at NON_GIVEN_NAME_PARTICLES fails the new value assertion while the pre-existing name assertion passes it (the bound set is identical either way); restoring the sys.modules subscript fails the new dir() test on all four shims.

Six further items were already fixed since the issue was filed — verified rather than assumed, and listed in the commit message.

The docs sweep

Three stale behavior claims, two of them broken by #296 in this same cycle — which is the shape the sweep exists for, since prose describes behavior in words and grepping the changed symbol finds none of it.

site	claim	measured
concepts.rst	Dr. is "a title before a name and a suffix after it"	"John Smith Dr." → family 'Dr.'
usage.rst	post-comma PhD lands in title	"田中さん, PhD" → suffix 'さん, PhD'
_group.py	Reads: "Policy.extra_suffix_delimiters — no other Policy field"	reads policy.lenient_comma_suffixes

concepts.rst's was the worked example for why the fields are really pre- and post-nominal; replaced with Lt., one of the twelve deliberate duals #296 kept. _group.py's undeclared read arrived with the #430/#432 review round (986e9d6) and the header was never updated — a false universal in a Reads: block is worse than an omission, being the thing the next person checks against instead of the code.

Sites verified accurate and left alone are listed in that commit, so the sweep is on record as having covered them.

Verification

• 6,196 tests pass, 156 skipped, 9 xfailed
• mypy: no issues in 109 source files; ruff: clean
• sphinx-build -b html clean; -b doctest 108 doctests, 0 failures
• Differential gate exits 0 at all three baselines, unexplained: 0 — 1104 names, 226 / 198 / 106 intentional diffs at 1.4.0 / 2.0.0 / 2.1.0
• No open Dependabot PRs on uv.lock

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.66%. Comparing base (865e6bf) to head (fe1bccb).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #477   +/-   ##
=======================================
  Coverage   98.66%   98.66%           
=======================================
  Files          45       45           
  Lines        3214     3218    +4     
=======================================
+ Hits         3171     3175    +4     
  Misses         43       43           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.