Share group's connective-join loop with the comma decision's title chain #617

#613 (PR #616) moved rules.md#C1's comma decision to the head of group (_pipeline/_comma.py). The decision reads the part after the comma before any join, but one join has to happen first: rules.md#P3's connective binding title words into one title (John Smith, Mr. and Mrs.). _comma._title_chains makes that join with its own copy of group's connective loop (_group.py, the P3 chain inside _group_segment).

The two copies already differ:

• Group merges a run of consecutive connectives before joining; _title_chains doesn't.
• Group skips a single-letter connective only in short names (total < 4); _title_chains always skips one.
• Group derives a "prefix" kind as well as "title"; _title_chains derives "title" only, which is all a postnominal part can hold.

Cost so far:

• _title_chains needed its own fix for a quadratic slice rebuild, found in Decide C1 once after classify and bind the comma part; fold P6 into assign (#613) #616's review, and its own _PREFIXED_SHAPES clock row (title_chain).
• The first difference shows on garbage input. Prof. Smith, and and Dr. reads title 'and and Dr.', family 'Prof. Smith'. b547c24 read title 'Prof. and and Dr.', family 'Smith'.

Proposal: extract group's connective-join loop into one helper, parameterized by:

• which kinds a join derives (title only, or title and prefix);
• the single-letter policy.

decide calls it title-only; group calls it as now. This is mechanisms.md#ONE-PREDICATE-PER-QUESTION.

Effort and risk:

• The loop leans on closures inside _group_segment (conj, title, prefix, merge) and on the frozen set, so the extraction needs care.
• Benchmark rows to re-check: title_chain, plus _SHAPES' conjunctions.
• The target is no change in output except the double-connective garbage case. That one should be decided explicitly and recorded in decisions.md.

Related: #613, #616, #611.