fix(maiden): a marker stops the particle chain that swallowed it

[derek73]

Closes #399.

Ursula von der Leyen geb. Albrecht lost the maiden name that the same words one particle chain apart kept:

Ursula Leyen geb. Albrecht          ->  family='Leyen'          maiden='Albrecht'
Ursula von der Leyen geb. Albrecht  ->  family='von der Leyen'  maiden='Albrecht'   (was: family='von der Leyen geb. Albrecht', maiden='')

The fix

Grouping's prefix chain absorbs everything up to the next prefix or suffix piece, and the maiden marker is consumed after the chain merges — by which point there is no lone marker piece left to find. The chain now stops at a marker the way it already stops at a suffix.

Only a non-leading particle ever reached the marker, which is why a leading single particle always worked (P4 chains nothing) while a leading run of two did not — the second particle's own chain fired.

The stop is gated on the consumer actually taking

This is the substance of the change, and it came out of review rather than the first cut. Sharing one marker-piece predicate settles what a marker piece is; it does not settle whether the consumer takes, because the consumer also needs a non-suffix piece after the marker. Stopping without one left the marker standing as its own piece, and a lone trailing piece takes a role field:

Jane van der Berg née    ->  middle='van der Berg'  family='née'     <- ungated
                         ->  family='van der Berg née'               <- gated

The marker became the surname and the real surname was demoted — the defect the stop exists to prevent, one field over. The gated reading is also what M2 already said should happen: a marker with nothing after it "is just a word", exactly as Jones née → family='née'.

Three measured consequences of the ungated form, each closed by the gate:

• Jane van der Nee moved a bearer of the attested surname Nee out of the family. M1 already names Nee as a surname.
• Ursula von der Leyen geb. — an ordinary truncated record — reported the marker as the surname.
• St St née stopped reporting its particle-or-given fork. Group's emitter is guarded on the chain having merged something, and an ungated stop made it merge nothing for a different reason than the guard assumes: an A1 violation of the shape A tussenvoegsel attachment decides a particle-or-given fork and reports no ambiguity, which A1 requires #405 tracks, and worse than A tussenvoegsel attachment decides a particle-or-given fork and reports no ambiguity, which A1 requires #405's, because it removes a report callers already see.

Two shapes fixed without a second change

• Ursula von der Leyen geb. Albrecht loses the maiden name — a particle chain swallows the marker where a suffix stops it #399's open question. Under a family-first order the marker used to compete for the leftover given slot, so de la Cruz née Vega reported given='née'. Consumed and dropped, it never reaches the placement, and both family-first orders now agree. A marker the consumer declines still reaches it — that is M2's "just a word" reading, not a leak.
• The wholly-particle family. Jane de la née Jones → family='de la' with a non-empty base, via Should a particle that joins nothing still count as a particle? (Jong van der and Jong, van der already disagree) #404's R2 reading. A 305,364-parse sweep at review found no empty-base or partition violation either side of the change.

Limits, now stated rather than implicit

The bound reaches only a marker standing as a word of its own, so P3's connective join and P5's bound-given join each still absorb one first — Jane van der Berg née y Jones and van der Berg, abdul née Jones are unchanged. The deleted Accepted: clause covered a class of which this fixes one instance and leaves two; both survivors are back in M2 with examples.

Docs

• rules.md#M2: the stop and its gate are in the statement, ahead of the examples — the citation checker cuts a statement at its first example line, so a rule stated after them can never be cited from code.
• rules.md#P2: its statement enumerated the join's stops and named neither the marker nor the suffix. Both are named now, and its two code citations move with it.
• decisions.md#M2: the false invariant is corrected; the quoted removal is verbatim so git log -S finds it; and the recorded reason for preferring the stop over a stage reorder is rewritten, because the example it cited does not discriminate — Anna Müller geb. von der Berg is identical before Ursula von der Leyen geb. Albrecht loses the maiden name — a particle chain swallows the marker where a suffix stops it #399, after it, and under a reconstructed reorder.
• docs/release_log.rst gets its Behavior Changes bullet, which the first commit omitted.

Verification

Tests carry the whole verification. No corpus name has a particle ahead of a marker, so the differential gate is blind to this class and its clean exit says nothing about the change — all three baselines exit clean (1.4.0, 2.0.0, 2.1.0), the 1.4.0 baseline verified genuine rather than the mislabelled cached tree, over 751 names unchanged.

A synthetic Jane van der Berg 旧姓 Jones row was added to reach the CJK corpus builder — the only route a case row has into any corpus — and then removed. It used that corpus's provenance as a back door, and it pinned nothing: measured with the shipped lexicon it produces the same segments and pieces as the Latin spelling, since script_segment does not split at the boundary. #413 is the honest fix — the issues harvester matches only quoted names while this tracker writes them in backticks, so it misses 104, including Ursula von der Leyen geb. Albrecht from #399's own title.

20 case rows plus two piece-level tests in test_group.py. The piece level matters because that is where the two halves can disagree without any field looking wrong enough to fail — the misaligned version passed the entire suite. Removing the gate now fails 6 tests; dropping the marker predicate's lone-piece guard, previously an unkilled mutant, now fails 2.

Full suite green: 4293 passed. ruff and mypy clean, _group.py and _post_rules.py at 100% statement and branch coverage.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.55%. Comparing base (c43d0a3) to head (e35c870).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #409   +/-   ##
=======================================
  Coverage   98.55%   98.55%           
=======================================
  Files          44       44           
  Lines        2977     2981    +4     
=======================================
+ Hits         2934     2938    +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.