Task 10 is checked with the honest state written beside it: branch → PR #48 →
CI green in 1m00s (run 30389174367). The squash merge is NOT done and is not
dev-story's to do — code-review comes between here and the merge, and `done` is
the MERGE's business (5.1 established this; 5.2 and 5.2b confirmed it).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…e story turns out to have a green exit

Three parallel layers. NO behavioural defect was found by any of them; the
Acceptance Auditor reproduced SIX of the seven mutations and not one observed
red set differed from the record. What fell is documentation truth, in a story
built to guard exactly that.

THE DECISION. `all()`'s witness stops the build on a new variant (error[E0004])
but does NOT force the variant into the list. Measured by RUNNING the lazy
repair: a bare `Self::NewThing => {}` arm leaves `all()` returning 2 of 3 while
the suite reports 90 passed. Guy's call — keep the mechanism, state its limit
truthfully, register the residue with owner 5.14. Two closures were built and
measured first and both rejected:

  · an `ordinal()`/`slot()` witness is GREEN with the variant missing (91
    passed) — it closes nothing, because `slot()` is never called on a variant
    that is not already in `all()`. This was the reviewer's own first
    recommendation, withdrawn on measurement rather than on argument;
  · a single-source `macro_rules!` works (`all().len() = 3`) but violates AC1's
    literal `[Self; 2]` return type, and makes adding a variant frictionless —
    the opposite of AC1b, which calls a third variant a FINDING.

`all()`'s doc now states the guarantee and its limit, and the witness carries a
comment naming the wrong repair for whoever arrives on the E0004.

ALL THREE LAYERS CONVERGED on a false row count in the enum's own guard
paragraph — "the four D13 rows that produce none" when it is two. Inherited from
an AC1b that contradicts itself four lines apart; the dev copied the wrong half
instead of raising it as a finding.

Also patched: `all()`'s doc claiming a new variant breaks it "twice" when the two
errors are alternatives along one repair path; the variant-set test's doc
asserting a guard it does not have; a re-count the story pre-named and skipped
(`Expectation::MustAbstain` 7 → 8, the eighth being this story's own test);
AC1's `Given` clause, refuted by the story's own M4 and traced back to the 4.6a
register bullet, now struck there beside a SECOND false clause that bullet's
first closure had announced but never struck; the cross-vocabulary test widened
from 2 of 6 pairs to all six and RE-PROVEN red under M2; and eight smaller
count, antecedent and placement defects across four documents.

CLAUDE.md's Epic 5 line, stale since 5.2 and 5.2b merged, is fixed here rather
than deferred — the drift was inherited, the occasion is this push.

REVIEWER'S OWN DEFECT, recorded rather than quietly fixed: two layers with write
access ran against ONE working tree, and the Edge Case Hunter observed another
agent's mutation mid-pass. Nothing was corrupted — tree verified clean, md5
stable, gate re-run whole — but the next multi-layer review gives each writing
layer its own git worktree.

Test count UNCHANGED at 267 (135 + 90 + 42): the widened test is still one test,
so the number says nothing about these patches. The re-run red set does.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>