Decide C1 once after classify and bind the comma part; fold P6 into assign (#613)

[derek73]

Closes #613.

#611 found that rules.md#C1 had grown by patching. segment decided what the comma means before any word was tagged, so it had to predict later stages: group's particle chain (#562's pair) and assign's reading of a capitals-settled run. Assign then re-decided #296's positional read with a second count. P6 showed the same sign on a smaller scale: assign held back from #602's run exactly the particle tail that post_rules would attach a stage later.

What changes

C1 is decided once, after classify, and the part it decides is bound before any join (cfdd4775).

• A new module, _pipeline/_comma.py, runs at the head of group.
• It reads the part after the first comma once, with segment_suffix_reading, passing in the ambiguous words C1's count licenses.
• It then sets the roles of a postnominal part, so no later join can reach into it. This is mechanisms.md#READ-WITHOUT-THEN-BIND, its second instance after M2's Rethink the maiden-marker rule (M2): a marker stands after the current name and takes words up to the trailing suffix run #601.
• segment now outputs only NO_COMMA or FAMILY_COMMA.
• script_segment asks its own narrower question (_postnominal_behind_a_whole_name).
• Deleted: segment's decision block, classify's comma-run caps branch, assign's comma reading and second count, _vocab.run_word_fold, and the settled-run mirror test.

P6's comma site is decided in assign (47fb2b21).

• The attachment is decided in the pass that reads the given part, and particle_tail is asked once.
• Assign asks P1's fold condition itself.
• Nothing moves. This was checked on a 103,786-text P6 grid (four policies; roles, tags, reports in order, initials), on the harness, and on the gates. A reviewer confirmed it independently over 634,452 parses.

Decided (Derek), each measured first

1. A connective surname counts its words before the comma (Why does De La Cruz, Ed lose its given name? #575's count). Ortega y Gasset, Dr. now reads title 'Dr.', given 'Ortega', middle 'y', family 'Gasset'. PR Read a comma part a credential opens as the postnominal part; titles past the second comma (#603, #611) #615's Vega rows reverse.
2. Nothing after the decision unsettles what it bound.
   • Why does John Smith, PhD DO DO read given 'PhD' and family 'DO DO John Smith'? #562's pair report goes silent.
   • A credential part holding particles side by side stays the credential run (De La Cruz, DO DO DO reads suffix 'DO DO DO').
3. No particle run, connective join or bound given name reaches into a postnominal part.
   • John Smith, Ph. D. and Mary Jones reads suffix 'Ph. D. and Mary Jones'.
   • The one join kept is a title chain: John Smith, Mr. and Mrs..
4. Two changes around the positional read.
   • The positional read reports as the comma-less name does.
   • A surname with a suffix in it, before the comma, is the listing form's family: Smith Jr., Esq. reads family 'Smith'; Smith Jr., V reads given 'V'.

Accepted garbage-input costs, each pinned by a case row:

• John Smith, PhD and Dr.
• John Smith, Ph. Ma D.
• John Smith, Ma. (now read as John Smith, Ma is)

The CJK radar tier is settled in decisions.md: the name is never divided where decide reads a family comma, and an invariant test holds that.

Measured (2026-10-06, against b547c24)

• Gates: exit 0 at all five baselines. At 1.4.0, intentional diffs go 631 → 632; at 2.3.0, 285 → 286. The two new corpus names are C1 examples.
• Corpora: four names move.
  • Two lose Why does John Smith, PhD DO DO read given 'PhD' and family 'DO DO John Smith'? #562's report.
  • Ph. D., Jr. moves; it's garbage, and its watched row is re-recorded.
  • 마틴 킹, MA changes only in its report detail.
• Harness (80,423 texts × 3 policies): 3,631 texts move, 1,773 of them with a role change. Every default-policy role mover is in a named class. The grid, the classifier order and the counts are written out in decisions.md#C1.
• Frames:

Name	b547c24	This PR
Smith, John	179	160
John Smith, PhD Jones	361	262
Smith, MD PhD	253	185
No-comma names, including the reference parse	—	unchanged

Review

Two rounds, each with four or two reviewers. Fixes are in bac1b510, 7dc7bfa8 and 505f1e8b:

• Dotted initials after a two-word surname read as a credential: García Márquez, G.J. has no given name #563's quadratic came back in decide's paired-initials scan while its guard counted a function the moved code no longer calls. The scan is fixed, and the guard is retargeted.
  • AGENTS.md now records the lesson: a frame guard counts one function.
• script_segment's stand-in drifted from decide.
  • The first draft asked only the vocabulary.
  • The first fix then skipped decide's numeral and policy conditions, which divided 김민준 박, MA V behind a family comma.
  • It is now held by test_the_division_never_outruns_the_comma_decision, with recorded negative controls.
• Comments and docs: about 45 stale comments fixed; the docs updated are rules.md C1/S2/W3, mechanisms.md, AGENTS.md and the release log.

PR review (d99b9845; four reviewers: code, comments, tests, silent failures):

• Empty head: an empty part before the comma made decide return early, so , PhD read given 'PhD' and , Jr. title 'Jr.'. Both read suffix again, as in 2.3.0.
• P6 behind an empty head: no family is named there, so H1 and M4 must decide before P6. That attachment is post_rules' again (, Mr. Jones vd).
• script_segment's stand-in: it counted a suffix before the comma as a unit (김민준 Jr., PhD). The invariant test now compares against decide on the undivided name.
• Three hidden quadratics: _title_chains, the paired-initials set copy, and the core-list filter, none visible to the frame guards. Each now has a calibrated _PREFIXED_SHAPES clock row.

/simplify pass (3320d700, ed0a2e3e):

• No behavior change in the first commit:
  • decide reads the part once;
  • decide empties the segment of a part it binds, so group's filter is gone;
  • script_segment asks once instead of twice, and accented Latin comma names are back under master's cost.
• script_segment now asks _comma.decide itself (decided with Derek) instead of a hand copy of the comma decision. The copy had drifted four times in review.
  • Names divide exactly where decide reads a suffix comma, and test_the_division_never_outruns_the_comma_decision asserts both directions.
  • Two tolerated CJK classes move, both toward consistency with decide: Dr 김민준, MA divides as Dr 김민준 does, and John Smith, Dr. J.씨 reads as John Smith, J.씨 does.

Not in scope: #614 (S2's rethink).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.98%. Comparing base (b547c24) to head (ed0a2e3).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #616      +/-   ##
==========================================
- Coverage   98.98%   98.98%   -0.01%     
==========================================
  Files          45       46       +1     
  Lines        4255     4352      +97     
==========================================
+ Hits         4212     4308      +96     
- Misses         43       44       +1     

☔ 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.