Two stages hand-copy each other's trailing-word logic after #601/#602; share one predicate each

[derek73]

Why. The /simplify pass on #609 found two places where one stage re-derives a decision another stage owns. Each pair agrees today. Nothing pins the agreement, which is the drift class mechanisms.md#ONE-PREDICATE-PER-QUESTION names: each site keeps passing its own tests while they disagree about an input neither covers. #609 left both alone because fixing them reaches outside its diff.

1. P6's particle tail, computed in assign and in post_rules.

• _assign.py's family-comma given-part run (Should a name word after a credential join the suffix? John Smith PhD Jones reads middle Smith PhD #602) works out p6_lo:p6_hi, the particles ending the part behind any post-nominals. It leaves those to P6, so the run doesn't absorb a word P6 will attach to the family (Smith, John PhD de).
• _post_rules.py's P6 attachment finds the same tail its own way. It steps back past pieces with no name role, then back over particle pieces, stopping at a lone ambiguous member whose role is SUFFIX.
• The two stop differently. Assign tests tags, post_rules tests roles plus AMBIGUOUS_ACRONYM_TAG. They agree only because the run has already given every non-particle word and every class member a SUFFIX or TITLE role. A change to either walk, or to the run's roles, could split them silently.
• Proposed: one _pieces predicate for "the P6 tail of this part", called by both stages, each passing its own test for a word that holds no name. Cost: probably one frame on every family-comma parse, so _CALL_BASELINE would be re-recorded. Measure before deciding.

2. The maiden take's candidate filter copies the suffix peel's admissions.

• _group._maiden_take decides which words at the end of a clause might belong to the trailing run (_clause_tail_word plus an inline numeral-shape clause). Then it lets tail_reading decide which of them it actually takes.
• That filter is a hand copy of peel_trailing's per-piece admissions (suffix piece, end-of-walk numeral against the word before it, ambiguous member) and of H5's word test. Its contract is "a superset of what tail_reading can take". Nothing states or tests that contract. A new admission in the peel would be silently missing from the clause's view.
• Rethink the maiden-marker rule (M2): a marker stands after the current name and takes words up to the trailing suffix run #601's own implementation already hit this once: the prototype left out the numeral fork, so John née Jones Smith VI kept VI in the maiden name.
• Proposed: one may_trail predicate in _pieces, written beside peel_trailing and owning the numeral arm, plus a sweep test pinning that every word tail_reading takes is one may_trail admits. It touches maiden-clause parses only, so it costs no frames on ordinary names.

Done when: no behavior change. The case table, all five differential gates and an equivalence fuzz like #609's simplify check (every field, ambiguity and initials string over the corpora plus a random grid) are identical. Frame counts on ordinary names don't rise.

Related: #601, #602, #609.