fix(post_rules): a trailing particle attaches to the family under a declared family-first order (#365) - #466

The release note claims 2.2 settles what a declared name_order means for
Latin-script names. Derek asked whether that is documented anywhere --
specifically the trailing particle, "Jong Anke de". It was not.

customize.rst's name_order section covered the basic family-first read,
the comma override, FAMILY_FIRST_GIVEN_LAST for Vietnamese, and the
LEADING never-given particle that outranks the declared order. It said
nothing about either change this release actually made: #467's trailing
particle, or #395's bound on how far a leading run reaches. rules.md#P6
and decisions.md#P1 have both, but docs/design/ is contributor
documentation and is not in the Sphinx build at all -- source_suffix is
'.rst' and it is not in the toctree -- so no user reaches it from the
published docs.

Adds a subsection built on the rule Derek stated in the #466 postmortem:
a particle joins FORWARD, and backward joining is licensed only by a
family comma or a declared family-first order, both being someone
SAYING the family came first. That makes the two spellings one fact
rather than two behaviours to memorise:

    parse("Jong, Anke de").family              -> 'de Jong'
    family_first.parse("Jong Anke de").family  -> 'de Jong'

Then why FAMILY_FIRST is the only order it arises under (it is the only
one that puts a trailing piece in a MIDDLE), the reach bound with the
default-order contrast, and where the two family-first orders diverge.

Eight new doctests, all measured before being written; the suite is 244
doctests, 0 failures. The #467 and #395 bullets now point at
:doc:`customize` for the behaviour and at decisions.md for the
reasoning, rather than only at the contributor doc.

Co-Authored-By: Claude Opus 5 <[email protected]>

Ten findings from two review agents, all in dc17090. The three that
matter share one cause: the new entry re-derived arguments
decisions.md#P6 already makes, instead of citing them -- which is the
duplication this very PR exists to delete.

- The Declined block now cites decisions.md#P6 for where "216 of 216"
  came from (a sorted(particles)[:14] slice, already recorded there as
  axis 2 in its purest form) rather than inventing "six hand-picked
  particles", which contradicted it and halved the sampling error.
- It also stopped attributing #466's failure to do/mc/vd. #473's
  "ambiguous-particle half" is the 37 non-never-given particles, and
  decisions.md#P6 records that #466 died on a never-given test that
  excluded all 37, van among them. Different partition, different
  failure.
- The recompute recipe said "any spread of families and givens". With
  F or G drawn from PARTICLES it is 0 of 70, not 67 of 70
  ("Doe, Van de" vs "Doe Van de"), so a reviewer probing particle
  behavior would have reported P6 broken. Now bounded and measured.

Also: AGENTS.md's corpus caution loses its standing count -- both the
drifted and the "corrected" wording quoted figures decisions.md#P6 had
already superseded; axis 11 is no longer classed rules.md-specific,
which contradicted its own "consults no rule statement" and would have
had a reviewer skip it on a decisions.md-only change like this one;
axes 10/11 are no longer said to come from "later arcs" (#291
predates #386); the root pointer drops its numeral; docs-review.md
stops naming a section commit 9f1d46e deleted; and the agent's
description names AGENTS.md, which the axes bind and it omitted.

Refs #473

Co-Authored-By: Claude Opus 5 <[email protected]>

…ohibition

Three sites said a never-given particle "cannot be a name at all" or
"can never be a given name", and used that to justify the readings
built on it. It is false: NON_GIVEN_NAME_PARTICLES supplies the
reading position leaves open and forbids nothing. Measured, all 33
members report given == themselves when parsed bare.

The conclusions those sites draw are right; the reason given for them
was not, and it is load-bearing. A particle OPENING the name has the
rest of the name to join forward to -- that is evidence about the
writing rather than about the word, which is why no declared
name_order contradicts it. Stating it as a property of the vocabulary
invites the opposite inference for a particle ENDING the name, where
there is nothing to join forward to and the writing is all there is.

This is the premise derek73#466 was argued from, and the second correction of
this same over-claim: decisions.md#P6 already records dropping "no
particle is a name by itself" for the same reason.

Behavior descriptions that the rest of derek73#467 will change are left alone
here; this commit moves no parse.

The 1.0s wall-clock bound failed four times across derek73#466 and derek73#474, always
on CI's 3.12 job, always between 1.01s and 1.08s, with master re-running
green each time. Three local methods disagreed with CI and with each
other: uninstrumented ~1% branch-over-parent (inside the parent's own
spread), coverage-instrumented indistinguishable, and the whole
benchmark file under --cov made the branch FASTER. The same harness
measured origin/master at 94-96ms and again at 89-92ms twenty minutes
later. A check that cannot separate a 1% change from a busy runner does
not fail safe; it fails expensively.

Call counts are deterministic. Budgets: 470 for parse(), 512 for
HumanName(), set at 408.4 and 445.4 with 15% headroom. A loose 5s
backstop stays for what call counts cannot see -- a stage that gets
slower without calling anything more.

WHAT THE OLD BOUND CAUGHT IS RECORDED, NOT LOST. Against the v2.1.0 tag,
back to back on one machine: 79.1ms -> 97.0ms for 1000 facade parses,
+23%, and 386.4 -> 441.4 calls per parse. The bound was not too tight;
the parser had grown into it, and the clock could not say so with
enough confidence to act on.

The largest contributor is a correctness decision rather than waste:
derek73#439 extracted the piece-level predicates so group and assign would
share one implementation (mechanisms.md#ONE-PREDICATE-PER-QUESTION,
which exists because two copies of "is this a title piece" drifted and
caused bugs). That turned inline expressions into ~51 calls per parse
while _group shed 8. The leaves are two membership tests each; there is
no hotspot, and recovering the cost means inlining them back at hot
sites, reintroducing the drift derek73#439 removed. Not taken -- recorded in
decisions.md#parse-cost with the per-stage and per-module numbers.

One repair was tried and reverted: collapsing post_rules' three role-
index scans into one pass saved ONE call per parse, because those sites
only run when their rule fires. Recorded so it is not re-tried, and as
the reason the budget is the instrument -- a harness with +/-10% noise
cannot tell a one-call change from a fifty-call one.

The budget binds: 80 synthetic calls per parse trips both tests with
the actual number in the message.

Axes 1-10 compare a document to another document, to its examples, or
to the implementation, so none of them can find a rule that is simply
wrong -- the failure that backed out derek73#461 and closed derek73#466 unmerged.

The worked invariant: the family-comma writing under the default order
and the comma-less writing under FAMILY_FIRST must parse alike. It
holds for 67 of the 70 particles and fails for do, mc and vd, which is
rules.md#P6's contested-membership set recovered from behavior alone.

Also corrects the axis count in three places and the claim that derek73#386
earned every axis, which axes 10 and 11 falsify.

Closes derek73#473

Co-Authored-By: Claude Opus 5 <[email protected]>

Form 4 under FAMILY_FIRST must parse as its form-2 comma twin, title
movement included. Generated over a small forked vocabulary (never-
given/ambiguous/multi-token particles, an ASCII and non-ASCII
family, doubled middles) rather than sampled, and vocabulary-blind on
purpose: decisions.md#P6 (the 2026-08-30 derek73#467 entry) measured 0 of
630 -> 603 of 630 over the whole 70-particle vocabulary, with the 27
failures being exactly the particle-and-suffix trio (vd, mc, do) -- a
real limit, not noise (rules.md#P6) -- and derek73#466, the rejected
predecessor, lost a given name outright on "van Berg Jan de" and
promoted a post-nominal into the given slot on "Berg Jan Jr. de".
decisions.md#two-input-invariants recomputes the same correspondence
a second way (a 6x6 family/given spread per particle) and supplies
this file's negative-control recipe. An earlier draft of this file
cited "0/216 -> 216/216", which decisions.md#P6 explicitly retracts
as a sliced-vocabulary artifact (`sorted(particles)[:14]`, which
happened to exclude the trio); this file's own count -- 96
parametrized pairs plus a full-vocabulary negative-control sweep --
is kept separate from and never confused with either retracted or
surviving number.

The comparison now also covers the ambiguity-kind set alongside the
seven role fields (the compare.py comparator shape), since
decisions.md#P6 (derek73#405) records that this design's own repaired
failure was exactly an ambiguity asymmetry between two writings of
one name; all 96 pairs agree on both. A second, non-parametrized test
sweeps the entire shipped particle vocabulary in the minimal pair and
asserts the disagreeing set is exactly {do, mc, vd} -- a recorded
literal per mechanisms.md#RECORDED-ROSTERS, guarded by a
mechanisms.md#VOCABULARY-OVERLAP-AS-PRECONDITION precondition that
the fixture family/given is not itself particle vocabulary -- and
resolves decisions.md#two-input-invariants' "Open: derek73#469" marker for
an executable form.

Form 5 is absent by construction -- no comma format puts the given
name last (derek73#469).

Co-Authored-By: Claude Fable 5 <[email protected]>