[derek73]

Closes #892 (point 4; points 1–3 landed in #914).

What goes

A search of the pipeline for code that reads an empty part before the comma found one site:

• P6's second attachment site in post_rules, which ran after H1 and M4 only when nothing stood before the comma (Rethink the suffix-comma rule (C1): decide the comma's structure once, after the words are tagged #613 kept it there).
• The and fam_pieces guard in _assign's comma site, which kept the main attachment off that head.
• The k == end empty-run guard in _inside_a_credential_run, which only the removed site could reach. Dropping it fails no test.

#890's floor for , Dr. fils is not empty-head code. It is rule C4's count of name words before the comma, and a one-word head meets it too (Sir Alexandre, fils).

What moves

Assign now attaches the trailing particle behind an empty head ahead of H1. H1 then finds a family and stands down. That moves the vd/mc tails behind a title that addresses by surname: 112 given parts per order on C15's invariant grid, and one snapshot name.

Name	Before	After
, Dr. Jones vd	family 'Jones', suffix 'vd'	given 'Jones', family 'vd', suffix-or-name
, Jones (MA) vd	suffix 'MA'	nickname 'MA' (its twin's reading; closes decisions/S13's "left as found")
, Dr. Jones vd, PhD (off the grid)	family 'Jones', suffix 'vd, PhD'	given 'Jones', family 'vd', suffix 'PhD', suffix-or-name
, de PhD vd (off the grid)	given 'de', family 'vd', suffix 'PhD', suffix-or-name	given 'de', suffix 'PhD vd' (the comma site's P1 guard, which the removed site lacked)

Each now reads as its twin Smith, … does, and as an ambiguous particle already read (, Dr. Jones van → family 'van'). None of the 10 real empty-head names in the local corpora holds a trailing particle.

Docs and tests

• rules/C15: the vd knock-on leaves the H1 Except; the Precedence bullet says P6 also wins over H1.
• rules/P6: gains a bullet for the empty head, and its Signal now reads "marked the family's place".
• decisions/C15 and P6: new entries with SUPERSEDES: lines, and the reversed parts are marked on the old heads.
• decisions/S13: an Elsewhere: line records the reading that moved.
• Invariant test: the H1 arm is simplified, with a count of 2889 (was 3225). The negative control was re-taken: still 1,098 failures.
• Field-sweep claim count: 1393 → 1396.
• Notes: the case row is renamed and its expectation reversed, and three case notes plus the particle_tail docstring no longer describe the removed site.

Review round (/pr-review-toolkit:review-pr)

Code, comment and test reviewers ran. Fixed in 73da45a:

• Two moved classes off the grid are now in the table above, in decisions/C15 and P6, and in case rows.
• Absolute rows pin , Dr. Jones vd Jr. and , Jane van Jr. (FAMILY_FIRST). The twin invariant now compares two readings of one site, so a defect there would move both sides.
• The H1 count is asserted per order (963 each).
• Comments: the test_state.py ownership comment is fixed, the rules/P6 quote is now verbatim, _inside_a_credential_run states its precondition, and the guard row was renamed.
• Lesson added to nameparser/_pipeline/AGENTS.md: when folding one site of a rule into another, compare the two sites' guards.

Not changed: the report detail "joins the family the comma named" still appears behind an empty head. That is wording only, and nameparser/AGENTS.md already records it.

Checks

The design-docs review ran on the change, and all six findings are fixed: stale mechanism notes, the decisions edit-kind format, and rules/P6's scope. The full suite, ruff and mypy pass. On the first run the serial benchmarks failed three wall-clock scaling ratios, the same flake seen on #914; all 43 passed on rerun.

🤖 Generated with Claude Code