Show the Policy spellings that type-check - #334

Four sites, two kinds of drift.

derek73#334 moved the guide to the Policy spellings that type-check, but only
the hand-written .rst. Four _policy.py docstrings teach the old ones
and autodoc renders them into modules.html, so the API reference and
the guide disagreed on the same page. Worse, derek73#337's new runtime
warning hands the user Policy(segment_scripts=()) -- the library
telling you, in a package that ships py.typed, to write an arg-type
error. A test now pins the offered spelling; the warning tests matched
on 'ja_segmenter' and never checked the actionable half of the
message.

Two stage headers understated their inputs. _extract declared only the
two delimiter policy fields while reading three Lexicon suffix fields
through _suffix_shaped, which is not a detail: that is the mechanism
letting a clause's content overrule the delimiter, and derek73#335 would
extend it. _post_rules omitted Policy.middle_as_family.

Checked mechanically rather than by eye: every stage's declared Reads
against the policy and lexicon attributes it actually touches.
_script_segment, _tokenize, _assign, _group and _classify were
already accurate.

_vocab has no Reads line by design, being a helper whose predicates
take vocabulary explicitly. Its 'no state' claim still holds.

…NTS.md files (#909)

* Move the ambiguity-emitter rules from AGENTS.md to nameparser/_pipeline/AGENTS.md

The Ambiguities bullet had grown to 1,488 words in the root file, which
every session loads. It is guidance for work inside the pipeline stages,
so it moves, unchanged in substance and split under subheadings, into a
nested AGENTS.md that loads when a session reads a stage. Root keeps the
principle and pointers to it and to mechanisms.md's contract entry; the
two test comments quoting "Pin the decision, not the vocabulary" now
name the new file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cut AGENTS.md's design-philosophy paragraph to its doctrine and an index of the script rules

The paragraph had reached 1,458 words, most of it the glued-honorific
peel's comma reach (#312, #319, #613) restated from code. Root keeps the
positional, never-detect-language doctrine, the script-itself exception
and its three-way test, and lists each behavior it licenses with the rule
that states it.

The facts no rule or decisions entry held move to decisions/W2 as a dated
entry, every parse re-run today: the reach and second-condition examples,
where the credential lands by spelling, why the run predicate is not the
token predicate pluralized, what lenient_comma_suffixes=False does, the
four drifts that made the peel ask _comma.decide, and that neither policy
opt-out reaches the peel. The #760 entry, which deferred two of those
facts to this paragraph, now says where they went. The opt-out example is
spelled segment_scripts=frozenset(), the denylist having refused the ()
spelling (#334).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Move the spelling sweep from AGENTS.md to a nameparser/AGENTS.md

The spelling sweep (1,326 words) is needed by any change to a word class,
vocabulary-only changes included, so it moves to a nested AGENTS.md at
nameparser/ rather than under _pipeline/: reading a vocab TOML loads it
too. The text moves unchanged in substance, split under a subheading per
class (numerals, credentials and titles, separators, full stops inside a
word, capitals, adjacent pairs). Root keeps the lead and a pointer; the
Workflow paragraph, which says where a missed shape is recorded, and the
two cases.py notes that cite the sweep now name the new file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Fix what the design-docs review found in the relocation

- Restore #337's `Policy(segment_scripts=())` in the validation bullet:
  the frozenset() replacement meant for the new opt-out example also hit
  this history line, which records () as the spelling that shipped wrong.
- Retarget the stale pointers: decisions/C14 (honorific-peel paragraph),
  _pieces.py and cases.py:2859 (the spelling sweep's #603 example and
  separator spacing).
- Root index: restore the katakana reason (its script settles nothing,
  not "foreign"), cite W1's interpunct Except for "never segments", name
  decisions/T3 as the nakaguro's history, keep 高橋+一郎.
- decisions/W2: the #760 entry's last pointer, and the facts the cut left
  only in usage.rst or nowhere (#319's failure mechanism, the pre-#319
  readings, why 博士/殿/君 are spaced-only), re-run 2026-10-27.
- nameparser/AGENTS.md: give the moved "That holds" sentence an antecedent.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cite #909 in decisions/W2's 2026-10-27 entry

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Record in AGENTS.md that a whole-file replace rewrites history lines (#909)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>