differential: compare initials() on roles-identical names (#484); fix the facade's dropped E/Y initial (#462) - #494

)

The gate compared the seven roles plus _ambiguities, so a render-layer
change with every role identical was invisible to it by construction.
Both surfaces now carry an _initials pseudo-field -- the facade's at
every baseline, the core's from 2.0 on -- and it enters a name's diff
only when the roles and ambiguity kinds agree: a role move drags its
initials with it, so that movement is the role diff's consequence, not
drift. Strict subset semantics were measured and rejected: they would
have put _initials onto a long tail of existing rules for no added
discrimination (the counts are in the decisions entry that lands with
this PR). The report prints _initials only under main()'s own verdict,
so an UNEXPLAINED block always pastes into a rule the validator accepts.

What it surfaced, now classified with fields = ["_initials"] rules:
the R2 all-particle readmission (27 names, three ledgers, shipped in
2.2.0 with a release note but no rule the gate could hold), #360's
'los' (1 name), and HumanName.initials() initialing each WORD of a
joined run where 1.4.0 initialed the run as one group -- 98 names at
1.4.0, changed in 2.0.0 and recorded nowhere until now. AGENTS.md's
release-note doctrine now says initials() is visible to the gate as a
floor, not a population.

The 14 names the facade's dropped E/Y initial moved (#462, fixed in
the previous commit) now move against the three 2.x baselines instead
and carry a fix(#462) rule in each, so the gate is green at all four
baselines at this commit.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

… R3 twin

The decisions entry for #484: why _initials enters a diff only when
the roles agree (strict semantics measured at 43 / 43 / 36 rules of
churn for no discrimination), the three clusters it surfaced with
their counts and tiers, why the per-word grouping is classified three
minors late, and the recompute recipe with its mutation check. R3
gains the #462 bullet. Two pointer comments so the next reader of the
case table's 'no initials column' note and the facade's 'not pinned'
comment knows the differential now looks, and one ledger header clause
reworded so #462 does not read as 2.2.0-era.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…ires

Four reviewers over the three commits. Documentation truth: the #462
names agree with 1.4.0 again since the fix, so two comments that said
they diff there now say what the connective rule's two exclusions
actually buy -- any future facade initials change on those shapes
surfaces UNEXPLAINED at 1.4.0 rather than being absorbed as per-word
grouping; R2, R3 and R4 carried the gate-cannot-see-initials limit in
the present tense and gain dated amendments; the README's standing
rule count is gone; _facade.py's "v1 parser.py:427 verbatim" says
"after ..., not verbatim", the loop no longer being a textual copy;
and the release bullet's John E Smith is a bare ASCII capital, since
_render._INITIAL is ^(\w\.|[A-Z])$ and an undotted Cyrillic capital is
not initial-shaped.

The initials-view entry's own measured claims re-derived on the
shipped tree: the 1.4.0 churn numerator (53 of 72, not 43), the
140 / 28 cluster totals dated to the pre-fix tree with the shipped
tree's 126 / 42 / 42 / 14 beside them, two illustrations replaced by
inputs that reproduce, the 14 given the population the bug could reach
(14 of 14, the rule's regex reaching four more in leading position).
The RECOMPUTE bullet splits into two recipes, because only one of them
runs off the gate: the classified totals are the sum of the
fields = ["_initials"] rules' heading counts, while the per-name split
and the pre-fix figures need the in-memory run.

One printer change, in both the UNEXPLAINED and the UNCLASSIFIED
(radar) block, which share _print_field_diffs: it shows the core's
initials movement when that differs from the facade's, the two being
independent implementations -- with a test for the differing pair and
one for the identical pair the print-once convention still covers.

The bound-given rule's ordering rationale is measured rather than
asserted: TWO corpus names carry both a bound word and a particle
chain, not three, and neither has an `_initials` diff -- the third
name the note counted shares with the Ph. D. rule.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Review round 2 conditioned the mixed-_initials check on the OTHER
names in `fields`, which made `["_initials", "_initials"]` pass where
it had been refused with an empty-list message. classify() reads
`fields` as a set, so a repeat matches nothing more; it is a
copy-paste slip that every check below would otherwise swallow, and
the `_initials` check in particular would read the pair as `_initials`
alone. Both validators now refuse it, with a row in each rejection
table.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>