`Smith, Ph. D. Jr.` puts the credential in `first` (regression from 1.4.0)

[derek73]

A space-split Ph. D. followed by another suffix lands in the given name:

HumanName("Smith, Ph. D. Jr.")
# 1.4.0:  last='Smith'  suffix='Ph. D.'  title='Jr.'
# 2.0.0:  last='Smith'  first='Ph. D.'   suffix='Jr.'   <- credential in first
# HEAD:   same as 2.0.0

Boundary

Measured. Regresses only when all three hold — one word before the comma, the space-split Ph. D. pair, and at least one more suffix after it:

input	1.4.0 first	2.x first
Smith, Ph. D. Jr.	''	'Ph. D.'	regresses
Smith, Ph. D. MD	''	'Ph. D.'	regresses
Smith, Ph. D. III	'III'	'Ph. D.'	regresses
Smith, Ph. D. Jr. MD	'MD'	'Ph. D.'	regresses
Smith, Ph. D.	'Smith'	''	fine — 2.x better
Smith, Ph.D. Jr.	'Ph.D.'	'Ph.D.'	parity (one token, no merge)
Smith, PhD Jr.	'Jr.'	''	fine — 2.x better
Smith Jones, Ph. D. Jr.	'Smith'	'Smith'	parity (two words → SUFFIX_COMMA)

Neighbouring shapes go the other way — 1.4.0 puts a credential in first and 2.x correctly routes it to suffix, the known fix(comma-family) improvement. This is the one row going backwards.

Why

Same root cause as #319, hitting a different victim.

segment() picks SUFFIX_COMMA only when suffixy(groups[1]) and len(groups[0]) > 1. With one word before the comma it falls through to FAMILY_COMMA even though the post-comma run is entirely suffix-shaped — so _assign reads that run as name text.

Why only the space-split pair: suffixy merges adjacent Ph./D. into one unit, and _group has a matching merge. For Smith, Ph. D. the merged piece is the lone post-comma piece, so the "lone post-comma piece routes to suffix/title" rule fires and it lands correctly. Add another suffix and it is no longer lone, that rule stops applying, and the merged piece falls through to the name-text reading.

#319 is this defect reaching the honorific peel. This is it reaching _assign. #319's fix does not address this — that one changes the peel's scope; _assign's reading of segments[1] is untouched.

Why the differential did not catch it

unexplained: 0 is a true statement about a corpus that cannot express the shape. corpus.jsonl has John Smith, Ph. D. — two words before the comma, which makes it SUFFIX_COMMA and works. There is no one-word-before-comma variant with a multi-token suffix run anywhere in the three corpora. Worth adding regardless of when this is fixed.

Milestone

Filed v2.2 rather than v2.1 on two grounds: it shipped in 2.0.0, so this is not a "stop it reaching users" situation; and the honest fix is at segment()'s len(groups[0]) > 1 rule, which is v1-parity machinery whose change needs its own differential pass and would not be a small addition to the 2.1 queue. Move it if you would rather take it sooner — the reproduction is tight and the boundary is known.

Found while investigating the Ph./D. merge duplication noted in #319's spec.