Word a report's field from the word's final role (#626)

[derek73]

Summary

Some ambiguity reports name the field their word was read into ("read as a given name"). Six of assign's reports took that field from the word's role at assign, but rules in post_rules can still move the word:

• H1 moves the name behind a title into the family.
• P1's family-first fold lays the rest of the name out by the order.
• P6 attaches a trailing name word to the family at the end of a name read family-first.
• Under opt-in policies, the patronymic rotations (O1, O2) and middle_as_family's fold (O3) move one too.

So the report could name a field the word doesn't end up in:

Input	Policy	Lands in	Report said
Attorney General of Minnesota	default	family	"read as a given name by convention" (since 2.3.0)
Kim Min Do	FAMILY_FIRST	family	"read as a middle name" (this cycle, after #289)
de Kim Ma	FAMILY_FIRST	given	"read as a middle name" (this cycle, after #289)
, Ma Dr.	default	family	"read as the given name" (this cycle, after #289)
Van Ivan Petrovich	East Slavic patronymic rule	family	"read as a given name" (since 2.3.0)

The report is still created where the parse makes its choice. Only the field word moves to assemble, the one place that knows final roles:

• An emitter that names a field writes its sentence up to the field, and passes the rest as PendingAmbiguity.field_tail. An empty string means the field ends the sentence.
• assemble joins the two halves around FIELD_READING for the first token's final role.
• The six emitters are H4's all-suffix name, title-or-name's join clause, O5's lone word, the trailing peel's picks, the leading ambiguous particle, and the report on the word right after a comma.
• A field report in a name emptied of content is dropped, rather than shipping half a sentence. No report reaches that case today.
• An import-time check confirms FIELD_READING covers every role.

One user-visible wording change beyond the field itself: the comma report's given-side text changes from "read as the given name" to "read as a given name", matching the other reports.

Measured (recipe in decisions.md#A1)

• Generated grid (53,517 texts, including commas with nothing before them and a trailing title, under all three orders; the opt-in movers are outside it): on master, 2,725 of 89,092 field claims named the wrong field. They came from three emitters: title-or-name's join clause (1,575), the trailing peel's picks (1,114) and the comma report (36). On this branch: 0.
• Fingerprint against master (differential corpora, case-table names and a comma grid, 13,090 names under six policies): no role, report kind, initials or rendering moves. The details that change:
  • 6,395 comma reports on 1,086 names: "the given name" becomes "a given name".
  • 85 comma reports on 15 names now name the family.
  • 40 title-or-name reports on 10 names now name the family.
• Opt-in policies: a review grid of 12,593 texts under 16 configurations (patronymic rules, middle_as_family, suffix delimiters, the caps and dotted switches, the RU, TR_AZ, ZH and JA locales) found 0 wrong claims on this branch, against 26 to 512 per configuration on master.
• Five differential gates: all exit 0. The reference frame-count pins are unchanged; names that carry a report are 2 frames cheaper than on master.

Tests

• test_a_report_names_the_field_its_word_lands_in (tests/v2/test_cases.py): every non-locale case row under all three orders. Rows that carry their own policy are checked under that policy only.
  • It recognises every phrasing on its _FIELD_CLAIMS list: assemble's, plus the fixed ones (the suffix run, the numeral fork, H4's "the name", P6's "joins the family"). Segment's "consumed as suffix" is not on it (Smith, Jr., Dr. Bart reports its last part as "consumed as suffix" when 'Dr.' is read as a title #629).
  • It and the existing R2 partition test walk one shared row list (_SWEPT_ROWS) and one cached parse per row and order.
  • Its negative control is 14 failures when run against master's code.
• test_the_field_sweep_sees_the_claims_it_checks: records how many claims the sweep checks (1155), so it cannot pass by checking nothing.
• New case rows, each label from git bisect against the wheels' readings:
  • Attorney General of Minnesota (parity, checked against the 1.4.0 wheel)
  • Kim Min Do (fix(#289))
  • de Kim Ma under FAMILY_FIRST (fix(#289), core-only)
  • , Ma Dr. (fix(#316))
  • Van Ivan Petrovich (East Slavic) and Van Ali Veli oglu (Turkic), both parity
  • Do Bishop Do under FAMILY_FIRST with middle_as_family (fix(#289), core-only)
• Unit tests for assemble's join, its handling of an emptied name, and which token's field it names. test_assign.py now reads details from the final parse, through run().

Docs

• rules.md#A1: new statement clause; interacts now lists H1, O1, O2, O3, P1 and P6.
• decisions.md#A1: new 2026-10-08 entry, including what the first draft and its reviews missed, an exact recipe for recomputing the numbers, and an Open: #629 line.
• mechanisms.md: AMBIGUITY-AT-THE-DECISION-SITE contract, with the Smith, Jr., Dr. Bart reports its last part as "consumed as suffix" when 'Dr.' is read as a title #629 exception noted.
• AGENTS.md: the decision-site bullet, plus a sweep-paragraph lesson that empty-head comma grids need a trailing title.
• Release log: one bullet, covering the two shipped cases (Attorney General of Minnesota and the patronymic Van).

Review

Three review rounds and a /simplify pass:

• 33da23a: fixes the first review: the sixth emitter (, Ma Dr.), P1's fold, and gaps in the guard and the docs.
• f49053b: fixes the second review: the opt-in movers (O1–O3), whose leading-particle report no test caught regressing until the new rows, and the docs' route to Smith, Jr., Dr. Bart reports its last part as "consumed as suffix" when 'Dr.' is read as a title #629.
• 22a7f97: fixes the third review: the recipe's details and A1's cross-references.
• aa48ff5: the /simplify pass: shared row list, shared helpers, no behavior change (fingerprint byte-identical).

A related report outside #626's class (segment's "consumed as suffix best-effort" after #603) is filed as #629.

Closes #626

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.01%. Comparing base (26cdb89) to head (aa48ff5).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #628   +/-   ##
=======================================
  Coverage   99.01%   99.01%           
=======================================
  Files          46       46           
  Lines        4481     4486    +5     
=======================================
+ Hits         4437     4442    +5     
  Misses         44       44           

☔ 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.