Parsing slows quadratically when titles and credentials alternate at the end of a name

[derek73]

Found by the _pipeline/ sweep for #553.

Symptom. Parse time for a name ending in alternating credentials and titles grows with the square of that tail's length. Each ratio is the time for 4× the input (py3.11, 2026-09-28, best of 5); 4× is linear.

shape	2.2.0	2.3.0	master
"John Smith " + "MA Prof. " * n	4.0×	10.5×	12.3×
"John Smith " + "PhD Prof. " * n	4.0×	9.7×	10.4×
"Jane Doe nee Smith " + "MA Prof. " * n	4.0×	3.9×	14.8×

The clause shape has not been released yet. Controls: "John Smith " + "Prof. " * 2n and "Doe, John " + "MA Prof. " * n both read 4.0×.

Mechanism. tail_reading's while True loop in _pipeline/_pieces.py runs once per title the H5 chain takes:

• Each pass splices out one trailing title, then calls peel_trailing over the whole of rest again.
• peel_trailing asks listed_lean/ambiguous_lean for every member of the run peeled so far, so n pairs cost n passes of O(n) work. At n=400, cProfile shows 402 peel_trailing calls and 80,200 ambiguous_lean calls.
• The loop also copies rest and prepends to titled on every pass.

The maiden clause's take in _group.py (_maiden_take / _release_reads_off) and trailing_start_past_titles run the same fixed point again, which multiplies the cost: 1,204 peel_trailing calls for the clause shape.

Scope.

1. Make the fixed point linear by carrying the peel forward between passes instead of redoing it from scratch. Readings must not move; the differential gate at every baseline is the check.
2. Add clock-guard rows to _PREFIXED_SHAPES in tests/v2/test_benchmark.py (added by Fix quadratic cost of a long given part after a family comma #557) for the no-comma shape and the clause shape. Record the negative control with repeated runs, as that table's comment does.

Frame guards might catch this one, since the cost is Python-level, but only on these input shapes, and no current shape reaches it.