Committed rules documentation: rules.md, decisions.md, mechanisms.md with enforcement tests - #381

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The denylist's first run caught its seed entry live in usage.rst and
release_log.rst -- Policy(segment_scripts=()) fails mypy; both sites
now hand out the frozenset() spelling the warning itself offers.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Five rules (H1, P1, O1-O3) with executable examples; decisions.md
gains P1/O1/O2 history; the 57-line rule-1b comment block becomes a
10-line citation. The adversarial probe falsified O2's 'exactly four
words' claim (a suffix-bearing five-word name rotates); the statement
now counts name words with titles, suffixes and nicknames set aside,
and the falsifying input is pinned as an example.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Old-numbering map in decisions.md; drop the gitignored-spec pointer
from rules_doc.py; PR citations on O1/O2 entries; runner rejects an
ungated [ja] annotation (the extra-absent CI split).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…es values

Citation ID grammar now admits bare and underscored mechanism slugs
(SPANS, FOLDED_TAG); the value grammar accepts one level of list
nesting for pieces= assertions, with a unit test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…divisions)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The spec's drafted litmus statement ('recognized by vocabulary, not
by written shape') was falsified by the live parser: the leading-
abbreviation shape heuristic is real ('Xyz. John Smith' takes the
title), and the abugida gap is its LIMIT, recorded as H2's Accepted
consequence. S2 records the bare-ambiguous-acronym behavior
('Jack Wei Ma' → suffix='Ma') as an Accepted consequence pending
#326-adjacent design work.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Ambiguities assertion form implemented in the runner; plan-deviation
phrases retargeted to decisions.md entries; 27 rules total.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…d reality

The batched adversarial probe falsified or sharpened: P2 (join runs
to the end), H2 (unbroken-run definition; comma carve-out), H3
(comma carve-out), C1 (lenient/strict prose), S2 (bare-ambiguous
words-to-spare guard), M1 (S1 wins), M2 (comma and particle
carve-outs), W1 (family-position precision), W2 (once; vocabulary),
W3 (rescoped to the comma doctrine; segmenter clause qualified), W4
(interpunct source-order carve-out), N3 (suffix-count artifact), T1
(wording). Falsifying inputs pinned as examples; all excerpts
updated in lockstep.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ase repair)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The denylist caught its own seed spelling being reintroduced in the
D1 decisions entry -- reworded to describe the bad spelling without
containing it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…on shapes

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

….0 promotion

Harvest measurements spot-checked at landing; the two flagged
uncertainties resolved against suffixes.py's own vetting block (the
per-90-surnames count became an invariant). The 3.0 shim-shaped list
is promoted from session memory into the 3-0-reevaluations section;
the memory file is now a pointer.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

All 35 gitignored-spec references (spec/conventions/amendment section
numbers) now point at mechanisms.md entries, rules, or decisions.md;
three new mechanism entries carry the promoted contracts
(LOCALE-PACKS-PURE-DATA, FACADE-CONTRACT, CONFIG-SHIM-SNAPSHOT).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…eview)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…y source)

Attribution: decision dates restored where entries carried release
dates (W1/W4/T3); D1 gains the actual decision (warn-not-raise, the
filterable-inertness rationale, the conditional ja hint); W2's #312
entry regains its provenance (gates inherited, not designed). New:
the codepoint-scope decision and half-flanked-interpunct Declined
under T3, the B7 sanctioned-extra Declined-with-measurement,
SELF-EXPIRING-GUARD and the extra-environment-split field note in
Verification shapes. STATE-OFFSET-CHANNELS contract admits
presence-only consumers (both interpunct readers, measured). New
verified examples: W1's segmentation ambiguity, W2 peeling across an
interpunct-divided name (the review's conjecture, confirmed live),
W4's nakaguro/interpunct order contrast.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…(primary source)

The 양/군 Excluded entry had attached the spaced-safety argument to
the glued exclusion — corrected from the vetting block itself, with
김지군 resolving the harvest's flagged uncertainty properly; 氏 gets
its real reason (王氏 historical form); 殿 regains the harm clause.
Four Declined dates corrected to git author dates (2026-08-05); the
_check_tree asymmetry diagnosis restored; D1 gains the autodoc
boundary fact. New: the Vietnamese middle_as_family
Declined-with-measurement, the fullwidth-colon M2 Accepted
consequence (#317 linked), the Japanese spaced-forms asymmetry in W
Background, Vietnamese facts in O Background, and the
CANONICAL-VOCABULARY-AT-THE-BOUNDARY mechanism.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…rimary source)

decisions.md#P2 claimed #367 made 'Sir de Mesnil' chain -- the
inverse of the truth (#367 REMOVED a chain; the family reading is
P1's fold), contradicting P1's own correct entry; both now agree and
the misfiled example moved to P1. The pieces= assertion form is
implemented (its first using rule arrived): new P4 states that a
leading particle chains nothing, with [[Sir],[de],[Mesnil]] and the
chained counterpart as structural examples. P1's family-first
boundary becomes the doc's first live deviates: marker (#368
superseded #359's working-as-intended sentence; intended
family='de la Vega', today family='Juan'), and its statement now
names #364 as open rather than pre-answering it. New: S3
(period-joined vocabulary -- the rule whose absence made a docstring
wrong four times), the particle-curation and esq Excluded blocks
(mirroring AGENTS' algebra, counts deliberately omitted), two P1
Declined entries (the strict-xfail back door; the not-a-title
predicate that st/do/freiherr break), the #365 two-orders-disagree
why, the deprecation-bridge template in 3-0-reevaluations,
VOCABULARY-OVERLAP-AS-PRECONDITION, and #372's standing caution on
LEDGER-RULE-SEPARATION.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…mary source)

LEDGER-RULE-SEPARATION's contract was falsified by measurement: a
strict fields-subset pair shares the regex tier in the live 1.4
ledger and file order decides between them (a pure reorder
reattributes seven names, caught only by _CROSS_RULE_WINNERS). The
contract now states the between-tiers/within-tier split, the stale
1639-figure is framed as the dated #372-as-filed measurement with
the #375/#376 correction, and 'never actually run' is fixed (run
twice in #375, failed). The whole dormancy arc (#328-#376) lands in
decisions.md: three shipped decisions ([[never]] monotone
exclusions, always-on dormant with three diagnoses, the 2.0.0
baseline gap) and four Declined-with-measurements. New:
CROSS-RULE-OUTCOME-PINS mechanism, the alternation-plus-sync-roster
second half on CURATED-VOCABULARY-ALTERNATION, the
narrowing-relocates-the-bug field note, and the roster list
completed on RECORDED-ROSTERS.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…source)

H1 gains the empty-family Accepted consequence the reviewer's own
worst error called for ('Sir John' → family='', executable); the H
Background carries the renunciate criterion (retains-a-family-name,
not religious-title — Rabbi Cohen measured correct); S2's ambiguity
flag is stated as firing in both readings with the assertable
example; decisions#H2's issue range narrowed to #343/#344. New
category: contested vocabulary memberships get Open blocks keyed to
the vocabulary set (#342/#346/#343/#344) and the rules.md preamble
names them as the third grep-invisible class. Excluded blocks gain
the non-CJK honorific-tails prohibition (जी, the 殿 criterion in
Devanagari) and ঠাকুর-is-Tagore under TITLES.
LOCALE-PACKS-PURE-DATA's contract states add-only — a pack can
never remove a base entry, which is why #342-class fixes must
change shipped vocabulary. W Background gains the script-member
no-op fact and the abugida divergence limit.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Decision dates (not release dates; git over narrative), harvest-
landing provenance discipline (uncertainties resolved from the
source, never by inference), and the primary-source review protocol
-- each practiced and review-corrected during PR #381, none
previously written down.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…(primary source)

The comma-suffix arc gains its true origin (#291 filed 2026-07-26)
and the missing earlier Declined -- splitting multi-word entries
into single words, declined on the name-swallowing measurements in
the issue body, which is why the seven entries were removed rather
than split. New Excluded block for the eight dead multi-word
entries (with the chargé/charge split and its precedent); 'born'
recorded as a never-shipped drafting invention; the never-given
ENCODING and freeze rationales (#293) land beside the particle
curation block. S Background states the per-word matching truth the
docs assumed everywhere. FACADE-CONTRACT closes the stays-warning-
free misreading; FOLDED_TAG gains its consumer-side strip
(Parser.revise, with the measured reordering hazard); two new
mechanisms from the session's fixes (WARN-AT-THE-CALLER,
LEGACY-STATE-SIGNATURE); C1's #296 marker gains the title-field
note so a future measurement doesn't misread it as stale.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ry source)

Two Background falsifications corrected: the H criterion was
conflating list membership (how the title ADDRESSES -- the Arabic
honorifics qualify while fully retaining family names) with the
separate #346 empty-family question (surname retention); the N
Background still called the typographic pairs caller-opt-in two
rounds after #273 shipped them as defaults. FACADE-CONTRACT's
'identical results' gains the RFC's own release-log-classified
carve-out (Kennedy-nickname measured). New decisions sections from
the arc: N2's cross-pair mis-extraction history with its Declined
offset-filter, ma-do, deviates-registry (option C -- why packs
carry no callables, now cross-linked from LOCALE-PACKS and
FORCE-A-DECISION), normalization-fold (the deliberate lower/casefold
asymmetry), render-default (the declined lossless née template),
and P3's Google Code provenance with the GC-namespace note on
legacy-rule-numbers. M1's canonicalization bullet gains its date,
origin, and the through-patches ruling; C1 gains the i/v blast
radius. Excluded: curly single quotes; the 2026-07-19
transliteration deferrals (Arabic, Greek, Ottoman, Hebrew,
sri/shri). N2 states the dangling-suppression filter A1 depends on;
R Background names the default view.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…rce)

Zero falsifications this round; all additive. New rule A2 (content-
free input parses to the empty name; born-empty ambiguities
survive, with the v1 contrast and the review-fix history).
normalization-fold gains the fixed-point requirement and the
_title_key corollary (the shipped-then-fixed silently-inert key).
New decisions: given-name-titles' twice-declined validation (the
'sir and dame' counterexample), N2's why-only-the-apostrophe
(Derek's double-quote distinguishing test), and the two-corpus
provenance (2026-07-24: tracker-harvested corpus because v1 test
banks are structurally blind to 2.0 additions). New mechanism:
AMBIGUITY-AT-THE-DECISION-SITE (every deciding stage carries an
emitter; a branch that changes nothing is not a decision). Two
verification field notes: guard-the-family parametrization, and
growth-guard calibration.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…rimary source)

P3 falsified and corrected: the Latin-capital initial veto on
single-letter connectives is now stated ('Jose E Maria Santos'
pinned) with the script asymmetry as an Accepted consequence —
Cyrillic blessed by #267's closure, the Latin half never separately
adjudicated (recorded as such). C section gains which commas count
(U+060C and U+FF0C segment, U+3001 does not — #265) with executable
examples. Dates disciplined per the convention: M2 and O2 now
distinguish design date from issue and landing; T1 credits #266 for
bidi existing at all; O4 gains its #270-body provenance and the
free-form-tuples Declined. W1 records the opt-in→default-on delta
from the filed proposal and the JMnedict Declined (staleness, no
frequency data, licensing). The stale Open-on-closed-#270 blocks in
O1/O2 become flagged status notes; nee's silently-resolved risk gets
its contrast entry; no_vowels' removal is recorded.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…(primary source)

normalization-fold's 'documented 1.4 deviation' inverted the
history: v1.4 has no casefold anywhere (measured against the tag) —
casefold is a classified 2.0 change, and the 1.x line was symmetric
BY DESIGN, which strengthens the do-not-fix-symmetric warning with
a real alternative rather than a hypothetical. The stale
empty_attribute_default 3-0 bullet (promoted from a memory already
three days stale at promotion) is replaced by the actual removal
record under removed-v1-surface (#44 origin, the z-Smith in-band-
signaling bug, substitute-before-format, tombstone-not-deletion,
readers-not-just-writers). New comparison-surface section: the
eq/hash trilemma, the cross-constants asymmetry, and two Declined
entries including the deliberately-issueless lenient-matching
resolved-as-no. The bridge discipline (warn in a released version
first) becomes a standing 3-0 rule; WARN-AT-THE-CALLER gains its
first measured instance; comparison joins rules.md's Not-in-scope
with a pointer.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Sweeping session memories against the new docs surfaced one more
falsification from the docs' own author: P2's 'join runs to the end'
is wrong for multi-group names -- a second particle starts a new
group and groups distribute positionally (the v1 Portuguese tests
pin it; 'Vincent van Gogh van Beethoven' → middle='van Gogh'), with
the #132 combined reading living in the surnames view (now
assertable and pinned). Promoted from memories: O1's opt-in
rationale (PR #154's Martin/Franklin/Benjamin finding), the
comma-suffix bundle's ship+repair amendment, the 2026-07-06
api-churn declines and the 3.0 config parking lot (contextvars
scoped override; Parser(constants) with #262's contract framing),
the bytes removal record, the property-key pickle skip's
dies-with-shim note, and the tags-are-ground-truth verification
field note (the twice-recurred 1.2.1-shim mix-up).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The gitignored specs' residue sweep landed: the retired #271/#272
slug taboo corrected (the selector was deleted by #333 AFTER the
review round that landed the taboo -- the tree moved under the
claim), the JMnedict licensing simplification fixed per the
amendment's own correction, and nine residue items: the 19-word
audit table with Derek's framing thesis (intent for unshipped v2.2
work whose only record was gitignored), the A6 phrase-peel note,
new rule P5 (bound given names -- closing a known extraction gap)
with its Excluded entries, the R2 all-particles core/facade
divergence recorded as unadjudicated, NFC-for-classification-only,
the initials-repertoire declines (#320), the script-table placement
reversal with the 〆 rider, Turkic orthography (dotless ı), the
declined script-resolved ambiguity kinds, and the bridge-fairness
converse (a warning must wait when the replacement doesn't exist).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

… memory, again

All cited commit SHAs resolve with matching subjects; four entries
carried reviewer-remembered dates one day off the git author dates
(the round-10 session asserted 07-19 'per git author dates' for
commits git dates 07-18 -- same-day for the N2 arc, which
strengthens its own story), and P3's Google Code provenance is
2013, not 2014. Fourteen delta issue/PR citations all verify.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The N3 contradiction resolved by measurement (decisions.md and the
_assign docstring said 'one non-title piece'; the code counts one
piece IN TOTAL, titles included -- "'Smitty' Dr. Jones" proves it,
and rules.md's Accepted consequence was right all along). W1 gains
the interpunct division-suppression Accepted (vocabulary and
segmenter alike stand down on 安东尼·陈志明, per #298) and W3 stops
implying otherwise. The 19-word audit table recovers its lost 19th
word from the spec (se: drop from TITLES, the PE/SE post-nominal
argument). The ma-do entry credits the right guard for 'Jack Ma';
the Anh Do example gains its contested framing (#385); the A2
citation moves out of mid-sentence; _group's docstring lists P4.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>