Closes #401
Closes #421

Both issues are one defect in rule P5's bound-given join: it "needs a name word to spare" and joins "the word after it", and neither the reserve nor the join knew what a name word was. They now ask one question — will assign read this piece as a suffix? — and get one answer.

The change

Mechanism. In the bound-given block, the walk assign will make is computed once (pieces from the bound word onward, minus the merged-credential pieces assign drops), and with it the one verdict that position decides: whether the last piece in that walk is a roman numeral assign's fork reads as the suffix. reads_as_suffix(k) is then "a suffix piece, or that numeral". The reserve counts such pieces out (#401 — V carries both vocab:suffix and initial, and _is_suffix_piece's initial veto is right for assign's middle-initial question, wrong for "does a family survive this join?"). The join declines the piece it would absorb when it reads as a suffix (#421), beside the marker decline it already had — which also stops merge()'s tag union from turning the joined piece into a suffix piece, the route by which the bound word landed in the suffix field.

The mirror is assign's fork condition for condition, over the walk assign actually makes. Four drafts that were not exact each lost a family or moved a name, and every one was caught in review and is pinned: the numeral is not the first name piece (not "the piece before it is not a title" — jr is title vocabulary too); "last" and "the piece before" are read over the pieces assign keeps; the piece before is taken as the join would leave it, bound word first (abdul J. V); and only the main walk has the fork — under the post-comma LENIENT reserve, suffix pieces alone count, as always. The fork itself — is_trailing_numeral_suffix with _ROMAN — moved from _assign.py to _vocab.py so both stages share one definition; assign calls it with the same arguments, and the regex-sync pin follows it. The #369 licence's own suffix clause is gone, subsumed.

Provenance notes

• abdul Smith V loses the family name — the bound-given join counts an initial-shaped suffix as a name word to spare #401 is shipped-since-1.x, as the issue said. A draft of this branch claimed "byte-identical to 1.4.0, a 2.0 regression" because the gate printed EXPLAINED NOTHING for the 1.4.0 rule. The name had diffed (1.4.0: first='abdul Smith', last='V'); the rule's fields lacked suffix, so the fields-only catch-all absorbed it behind its capped printout. Fixed — the catch-all's count drops 16 → 15 — and recorded in the decisions entry and the 1.4.0 ledger comment as the inert-measurement hazard it is.
• The credential shape's 1.4.0 reading was given abdul Smith (v1's fix_phd extracted the credential before parsing, so the join took Smith), not "given abdul, suffix Ph. D." as the issue's fix section assumed. The fix restores the suffix field and deliberately not the given: the join takes the word after it or nothing, rather than reaching past a credential. A judgement, recorded as one — flagging it in case you'd rather the join skip an inner suffix.
• abdul Smith V Ph. D. is pinned at the field level only — a trailing split Ph. D. is the shape the 1.4.0 ledger's [[never]] entry refuses to explain any diff on, and v1's reading of that name carried the abdul Smith V loses the family name — the bound-given join counts an initial-shaped suffix as a name word to spare #401 family loss, so it stays out of the rules corpus rather than the never-entry being narrowed for one spelling.

Verification

• Suite: 5235 passed, 212 skipped, 10 xfailed; ruff and mypy clean. Nine group tests and the parser pins written RED first; the bound_given_join_sees_only_the_surviving_name case row re-pins to given abd, middle Jr Smith as the juan y garcia nee jones loses the family name — P3's word count includes the words the maiden name takes away #418 entry anticipated.
• Differential gate exits 0 at 1.4.0, 2.0.0, 2.1.0. No differential corpus name had either shape; the six new rules.md examples carry them into the rules corpus, where fix(#401) explains abdul Smith V and abdul Smith Jr V at every baseline (three fields at 1.4.0, which did not even read the V as a suffix; two at 2.x), and the two fix(#421) rules explain abdul Jr Smith Berg and abdul Ph. D. Smith Berg at every baseline.
• A new bound_given benchmark shape: the first draft of the walk was rebuilt per piece — 13.2× growth for 4× the input against the 6.0 bound, found by the PR review because no benchmark unit led with a bound word. Computed once it reads 4.2×, inside the clean column.
• Reviews: design-docs-reviewer, then code-reviewer × 2, pr-test-analyzer, comment-analyzer. Everything they found is addressed in commits 3–5 except the two follow-ups below.

Commits

1. fix(group) — the predicate, the move to _vocab, tests, case row, regex-sync pin, S2's citing module. Green and gate-clean on its own.
2. docs(design) — rules.md#P5, decisions.md#P5, regenerated corpus_rules.jsonl, ledgers at all three baselines, guard pins, two release bullets.
3. perf(group) — the walk computed once; the bound_given benchmark shape (RED at 13.2, GREEN at 4.2).
4. test(parser) — field-level pins for abdul J. V, abdul Smith V Ph. D., Berg, abdul Jr Smith, and (abd, abu) × (I, X); the case row's note corrected.
5. docs(design) — the acronym-fork claim narrowed; the 2.x fix(#401) rules trimmed to the two fields their diffs have; review-found comment rot.
6. refactor(group) — the /simplify pass: the mechanism said once with the rule cited, Lexicon.add reused, a redundant assertion dropped, ledger comments reduced to ledger mechanics, decisions pointed at John van der Berg V reads family van der Berg V — the particle chain and the maiden walk take a trailing numeral as a name word #424/abdul Smith Jr Ma loses the family name — P5's reserve does not mirror S2's bare-acronym fork #425.

Follow-ups, filed

• abdul Smith Jr Ma loses the family name — P5's reserve does not mirror S2's bare-acronym fork #425 — S2's bare-ambiguous-acronym fork is not mirrored by the reserve. abdul Smith Jr Ma → given abdul Smith, family '' where John Smith Jr Ma keeps family Smith; 132 such shapes, pre-existing. Framed, on the altitude review's advice, as sharing assign's peel over a post-join view rather than mirroring a second fork by hand.
• John van der Berg V reads family van der Berg V — the particle chain and the maiden walk take a trailing numeral as a name word #424 — The P2 particle chain and the maiden walk take a trailing bare numeral as a name word (John van der Berg V → family van der Berg V; John née Jones Smith V → maiden Jones Smith V). Same root cause as abdul Smith V loses the family name — the bound-given join counts an initial-shaped suffix as a name word to spare #401 at two more suffix() sites.

A /simplify pass (reuse, simplification, efficiency, altitude) ran after the review round: efficiency found nothing (the new work is ~1 µs on a 55 µs parse); altitude judged the mirror right for a minor release and supplied the #425 framing; the reuse and simplification findings — comment bloat, a hand-rolled Lexicon.add, a redundant assertion, ledger comments carrying parser mechanism — are commit 6.

🤖 Generated with Claude Code