Decouple the chained-emitter tests from the shipped particle vocabulary

[derek73]

Groundwork for #360, which we didn't want blocked on a vocabulary/test coupling. It turned out the coupling was never real, and the guard that enforced it was asserting the wrong set.

Two commits: the sweep, and the fixes from a review that found the sweep half-done. Both are described below.

The guard asserted the wrong intersection

test_the_chained_emitter_is_still_reachable required a lead word in titles & particles_ambiguous. What the emitter actually requires, measured against "Freiherr von Richthofen" one membership at a time:

lexicon	chain	fork reported
shipped	yes	on von, from _group
freiherr out of particles	no — given='von'	on von, from _assign
freiherr out of titles	yes — family='von Richthofen'	on Freiherr, from _assign
von out of particles_ambiguous	yes	none

So the two memberships buy different things: particles is what produces the chain, titles is what the emitter's all(title(x) for x in range(k)) guard needs to report it — the merge() sits outside that guard. Both intersections are ['do','freiherr','st'] today, which is why the wrong set read as correct.

#360 cannot empty the right one. Curation moves words between the may-be-given and never-given halves; both are subsets of particles, so particle membership survives either way. The emitter was never at risk.

The coupling was unnecessary anyway

Reachability is a property of the emitter, not of the shipped word lists. A caller can configure the overlap themselves — Lexicon asserts no invariant against it and construction warns about nothing:

Zzt van Johnson, 'zzt' as title only                  -> given='van'  family='Johnson'      (_assign's half)
Zzt van Johnson, 'zzt' in titles+particles            -> given=''     family='van Johnson'  (_group's half)

An empty shipped intersection would leave the emitter reachable and merely unexercised by default. The old guard's remediation advice — "remove the emitter rather than repointing them" — would have deleted live code on a false signal.

What changed

The emitter tests supply their memberships instead of borrowing them, split by the role each word plays:

• _TITLE_PARTICLES — the leading words, given titles + particles
• _CHAINED_PARTICLES — the chained words (von, van), given particles_ambiguous

test_the_chained_emitter_is_reachable_by_construction carries three controls, one per membership the claim rests on: drop the leading word's particles and the chain goes; drop the chained word's particles_ambiguous and the report goes; drop the leading word's particles_ambiguous and nothing moves. That third one is the correction made executable — it is the membership the old guard asserted, and it buys the emitter nothing.

tests/v2/test_properties.py carried the same borrow on a larger surface: its "Freiherr " lead feeds a sweep over every ambiguous particle, and if that lead goes transparent the _group shapes leave the sweep while it keeps passing. Supplied there too.

Shipped-vocabulary behavior stays pinned where it belongs — tests/v2/cases.py's Freiherr von Richthofen row. Note that row tracks the parse, not the memberships: moving freiherr between the particle halves leaves it green, which is correct.

Review round

A three-agent review (test coverage, comment accuracy, general) found the first commit half-done, and found the comment block claiming otherwise. All fixed in the second commit:

• the chained word was still borrowed — von/van came from the shipped lexicon, so moving them to never-given failed seven of these tests, the exact breakage the commit set out to prevent
• the fixture repeated the error — it added the leading word to all three sets, so nothing could contradict the wrong precondition; measured, that third membership is inert
• "drop it from either and there is no chain" was false for the titles half (see the table above)
• the new docstring opened with "both a title and an ambiguous particle" — verbatim the wrong set the commit exists to retire
• _overlap_parser's rationale cited 'Do St Johnson', which no test in the file parses

Verification

condition	result
full suite	3197 passed, 41 skipped, 11 xfailed
tests/v2/test_parser.py	91 passed, 1 xfailed
leading word out of titles+particles	only the cases.py tripwire fires
chained words moved to never-given	no test in this diff fails
_group's emitter suppressed	6 failed — still sensitive

Every simulation asserts it took effect before running, so none is a no-op measurement. ruff check and mypy clean.

Two tests outside this diff still borrow van's ambiguity — test_parsedname_repr_includes_ambiguities_line and test_leading_particle_detail_names_the_role_it_got. Left alone deliberately: they fail loudly with an obvious fix (repoint to another ambiguous particle), which is not the failure mode this PR is about.

Groundwork for #360.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.51%. Comparing base (5320a42) to head (ef43b83).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #378   +/-   ##
=======================================
  Coverage   98.51%   98.51%           
=======================================
  Files          44       44           
  Lines        2895     2895           
=======================================
  Hits         2852     2852           
  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.