Remove the empty head's own P6 site (#892)

docs/design/decisions/C15-comma-with-nothing-before-it.md

@@ -4,16 +4,19 @@
 
 - A comma with nothing before it is a family comma naming no surname, the part after it read as the given part after a family comma reads, held by an invariant over two inputs rather than by a sweep shape: `, X` reads as `Smith, X` reads, less the surname, its fields and ambiguity kinds alike; a particle P6 attaches is then the family alone (`, Jane van` → family 'van') — 2026-10-27 #892.
 - H1 is the one exception: a title and one name word, or a bound pair standing for it, read as the family (`, Dr. John` → family 'John', `, Dr. abdul Rahman` → family 'abdul Rahman'), the comma naming no surname to decide the field by, as a comma with nothing but titles after it names no given part — 2026-10-27 (Derek).
+- No code is written for the empty head alone: P6 attaches behind it as behind any family comma, ahead of H1, so a particle ending the name leaves H1 two name words and the exception does not apply (`, Dr. Jones vd` → given 'Jones', family 'vd') — 2026-10-27 #892 (Derek).
 
-Open: whether the code written for the empty head alone (#890's floor for `, Dr. fils`, the P6 attachment site in post_rules) can go, the question #892's fourth point asks; the invariant answers it by mutation, one site at a time.
+Open: none in this file.
 
 ## Log
 
-- 2026-10-27 #892 — A COMMA WITH NOTHING BEFORE IT IS A FAMILY COMMA NAMING NO SURNAME. The shape had no rule: clauses in rules/H8, rules/O6 and rules/M2 and review findings in decisions/C1, S10, A1 and H5 each decided it locally, and nameparser/AGENTS.md made every comma change sweep it. Real names almost never take it: 10 of about 6 million local-corpus names (FEC 4, tse 3, PSC 2, tw-donors 1, the last opening with a fullwidth comma, `，邱詩涵`, which reads as `, 邱詩涵` does), the issue's count of 2026-10-25 missing the fullwidth one; recounted 2026-10-27 by scanning each local corpus's names for one opening, after spaces, with `,` or `，`. The alternative, dropping the comma so `, John Smith` reads as `John Smith`, was declined: the shape's one source is an empty surname field in listing data (#613's review, `, PhD` and `, Jr.`), and reading it as a whole name contradicts that source.
+- 2026-10-27 #892 (its `vd` and `mc` bullet superseded by 2026-10-27 #892 (Derek), below) — A COMMA WITH NOTHING BEFORE IT IS A FAMILY COMMA NAMING NO SURNAME. The shape had no rule: clauses in rules/H8, rules/O6 and rules/M2 and review findings in decisions/C1, S10, A1 and H5 each decided it locally, and nameparser/AGENTS.md made every comma change sweep it. Real names almost never take it: 10 of about 6 million local-corpus names (FEC 4, tse 3, PSC 2, tw-donors 1, the last opening with a fullwidth comma, `，邱詩涵`, which reads as `, 邱詩涵` does), the issue's count of 2026-10-25 missing the fullwidth one; recounted 2026-10-27 by scanning each local corpus's names for one opening, after spaces, with `,` or `，`. The alternative, dropping the comma so `, John Smith` reads as `John Smith`, was declined: the shape's one source is an empty surname field in listing data (#613's review, `, PhD` and `, Jr.`), and reading it as a whole name contradicts that source.
   - Measured on master (8a65ee18) before any change, over the grid `tests/v2/test_properties.py::test_an_empty_head_reads_as_its_family_comma_twin` builds (10 title runs × 32 bodies × 11 trailing words, 3,519 given parts under each of three orders): every given part reads as its twin but 1,075 per order, all of them H1's, so the rule landed with no code change. A first comparison's eight differing given parts holding `MA`, among them `, MA`, `, (Bob) MA`, `, J. MA` and `, JOHN SMITH MA`, were the twin's doing, not the empty head's: a surname written `Smith` brings a case contrast the given part lacks, so `Smith, MA` reads suffix 'MA' by its capitals (rules/S6) while `, MA` and `SMITH, MA` both read given 'MA'. The test writes the twin's surname in the given part's case for that reason, and rules/C15's Accepted bullet records the reading.
   - The particle-and-suffix words `vd` and `mc` ending the name behind H1's family reading (`, Dr. Jones vd`) are the exception's knock-on, not a second one: the twin attaches the word to its surname (rules/P6), the empty head has no family to attach it to once H1 has made the lone word the family, and it reads as the suffix. Behind a given-name title H1 reads no family and P6 attaches as in the twin (`, Sir Jones vd` → family 'vd').
   - The exception is stated by its result, not by the comma-less name: `, Jane Mrs` reads family 'Jane' where `Jane Mrs` reads family 'Mrs' (a bare salutation is a title only after a family comma, rules/H8), and `, Dr. abdul Rahman` reads family 'abdul Rahman' where `Dr. abdul Rahman` reads given 'abdul', family 'Rahman'. A first draft compared against the comma-less name and the design-docs review found it false on 493 of the 1,075 H1 rows per order.
   - The invariant compares fields and ambiguity KINDS, never a report's detail: `, Sir Jones vd` reports that 'vd' "joins the family the comma named", a comma that names none, and passes. Report-detail grids keep the empty head (nameparser/AGENTS.md).
   - Recompute: `uv run pytest -n0 "tests/v2/test_properties.py::test_an_empty_head_reads_as_its_family_comma_twin"`; its docstring records the negative control.
 
 - 2026-10-27 (Derek) — H1 STAYS THE ONE EXCEPTION, stated as the comma deciding nothing. "Keep H1 as the one exception, stated as a comma with nothing before it not deciding the field." `, Dr. John` keeps family 'John', where its twin `Smith, Dr. John` reads given 'John': rules/H1 already says a comma with nothing but titles after it names no given part and so does not decide the field, and a comma with nothing before it names no family, the mirror case. Following the invariant instead would have read given 'John' for every such row, 1,075 per order on the grid; rules/H8's Accepted bullet records the trailing spellings' family reading (`, Jane Mrs`, and the period-marked one since H5, #316).
+
+- 2026-10-27 #892 (Derek) — THE EMPTY HEAD'S OWN CODE IS GONE, the question #892's fourth point asked. SUPERSEDES: the `vd` and `mc` bullet of 2026-10-27 #892, which read `, Dr. Jones vd` as suffix 'vd'. 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 after H1 and M4, with the guard in assign's comma site that kept the attachment from acting there (decisions/P6, #613). Switching both off moved 112 given parts per order on the invariant grid, every one a particle-and-suffix word `vd` or `mc` ending the name behind a title that addresses by surname (`, Dr. Jones vd`, `, Mr Jones mc`, `, Jones vd Mrs`), each from family 'Jones', suffix 'vd' to given 'Jones', family 'vd' with a `suffix-or-name` report, and moved them onto the twin: no row failed the invariant, and H1's rows fell from 1,075 to 963 per order. Off the grid the PR review found two more classes, each onto its twin too: the same tail behind a third comma part (`, Dr. Jones vd, PhD` → family 'vd', suffix 'PhD', where it read suffix 'vd, PhD'), and a never-given particle opening the part with no name word behind it (`, de PhD vd` → given 'de', suffix 'PhD vd', as `Smith, de PhD vd` and `, de vd` read, where the removed site, which lacked the comma site's P1 guard, read family 'vd', suffix 'PhD', reported). An ambiguous particle there already read so (`, Dr. Jones van` → family 'van'), and the site had kept the comma-less name's reading (`Dr. Jones vd`). Removed, with "ok, remove it". #890's floor for `, Dr. fils` is not the empty head's: it is rules/C4's count of name words before the comma, which a one-word head meets too (`Sir Alexandre, fils`). None of the 10 real names with the shape holds a trailing particle.