Add "abd" to BOUND_GIVEN_NAMES, and let a word in two vocabularies satisfy the join reserve

[derek73]

abd Allah Smith read given abd, middle Allah, where abdul Rahman Smith reads given abdul Rahman. The spellings that write the article as its own word — Abd Allah, Abd al-Rahman — match none of abdul/abdel/abdal, and the Arabic-script عبد has covered the same word since #269, so only the Latin side was short.

abd Allah Smith          ->  given='abd Allah'   family='Smith'      (was given='abd', middle='Allah')
Smith, abd Allah         ->  given='abd Allah'   family='Smith'
abd Rahman Ahmed Salem   ->  given='abd Rahman'  middle='Ahmed'      (pairwise, like abdul)

Why the wordlist alone doesn't do it

abd is also the postnominal ABD ("All But Dissertation") in SUFFIX_ACRONYMS. The join's reserve counted every piece except titles and suffixes — so the bound word's own piece didn't count toward the three it needed, and the rule declined silently on the common three-word shape. Adding the word and stopping there changes nothing:

                                  abd Allah Smith        Jane Smith, ABD
wordlist only                     unchanged (broken)     suffix='ABD'
+ move to acronyms-ambiguous      given='abd Allah'      given='ABD'  <- credential lost
+ drop from SUFFIX_ACRONYMS       given='abd Allah'      family='ABD' <- credential lost
+ fix the reserve count           given='abd Allah'      suffix='ABD'

The last row is this PR. The claimed piece now counts, which is what the reserve means: it asks whether enough other words are left to spare, and the piece the rule has already claimed as a name is not one of them to spare. Both readings survive — position tells them apart, so Jane Smith ABD, Jane Smith, ABD and Jane Smith A.B.D. all still read the credential as a suffix.

Same shape as #397

Dual set membership shrinking a count that gates a rule is exactly the Catalan i finding — there it shrinks the rootname count P3's single-letter carve-out tests. Second instance in two days. The general question (should membership in one set suppress a word from a count gating a rule keyed on another?) stays open on #397; this PR fixes only the instance in front of it, where the count was wrong on its own terms.

Verification

abd is the only shipped word in both sets — measured, not assumed — so nothing else can move. All 751 differential corpus names are byte-identical and the harness reports 0 unexplained at the 1.4.0 baseline. The corpus contains no abd name, so that is a blast-radius check, not evidence for the fix; the evidence is the tests.

Mutation-verified: reverting the count line alone puts abd Allah Smith back to given abd. The stage-level test in test_group.py uses a synthetic word in both vocabularies, so it pins the mechanism rather than the wordlist, and asserts the reserve still declines at two name words.

3522 tests pass, mypy clean. Per-word evidence comments added on both sides of the collision, in the #: style, so a future editor of either set sees the other.

Review round

Four reviewers ran against this branch. The code review could not break the change and proved the blast-radius claim rather than sampling it: the join's own guard requires a one-token piece, and a one-token piece can never carry a suffix ptag, so dual membership collapses to exactly the set intersection — {abd}. Spellings cannot sneak in either, since A.B.D. normalizes to a.b.d for the acronym test and never matches the bound word. Empirically: 0 diffs over the shipped corpora, and 0 over 124,535 constructed abd-free inputs across all three name orders.

What the reviews changed here:

• An assertion in the stage test was inert — dual Allah never reaches the reserve, an outer >= 3 gate short-circuiting first, so it passed even with the reserve broken. Replaced with dual Allah jr, which the tempting simpler repair (count every non-title piece) fails by joining and leaving no family name.
• The changed line is a suffix count and nothing combined a firing join with a suffix. Six such shapes reverted silently under mutation; they are now pinned, along with abd Allah Smith ABD — both readings in one string, which is the collision's whole thesis.
• Two comment claims were false: "Abd Allah" and "Abd al-Rahman" are two-word and do not change at all, and the stated rationale ("spellings that write the article as its own word") names the one spelling still not joined — abd al rahman smith, where al chains forward. That half of the original exclusion survives and is now recorded as a known limit.
• "Position tells them apart" was too strong: in the given slot of a family-comma name the credential still wins (Smith, Abd → suffix='Abd', no given name, no ambiguity emitted), where Smith, Abdul gives the given name.
• decisions.md still listed abd under Excluded seven lines below the entry adding it; the constant's own docstring stated the reserve rule this PR falsifies; عبد shipped in 2.0, not 2.1; and the suffixes.py comment named عبد as the dual member when it is abd.
• This is not a C-ii collision. C-ii settles a precedence by frequency where neither reading is a name. Here one reading is a name and both survive, decided by position — which is C-i's positional qualifier, the de row of that table rather than the vd row. Relabelled, because C-ii is cited elsewhere for a real precedence and the two must stay distinguishable.
• rules.md#P5 now states its reserve and its S2 precedence, the way P6 already does for the identical situation. P5 previously had no interacts: line at all.
• The release-log claim is reframed: 0 of 751 corpus names carry a dual-membership word, so byte-identity there tests the corpus. The real argument is arithmetic — for a non-suffix first piece the new expression reduces to the old one.

A pre-existing bug found in passing is filed as #401 (abdul Smith V loses the family name; the reserve counts an initial-shaped Roman numeral as a name word to spare). Not addressed here — it predates this PR and reproduces on master with abdul.

mechanisms.md gains one field note from the review method itself: re-exec'ing a module in-process rebinds its enums to new classes, so is comparisons against the original module's members answer wrong silently — that produced 3,150 phantom movers in one sweep before it was caught.

3531 tests pass, mypy clean, differential 0 unexplained.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.53%. Comparing base (55a7822) to head (268b47e).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #400   +/-   ##
=======================================
  Coverage   98.53%   98.53%           
=======================================
  Files          44       44           
  Lines        2942     2942           
=======================================
  Hits         2899     2899           
  Misses         43       43           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.