Test 3 asserted phrase presence over the whole requirement block, so a referent
could leave the sentence carrying the normative force and survive in the closing
recap. Execution measured that escape and the annex recorded it; this commit
closes it by strengthening the guard rather than softening pin 4a's claim.
Test 3 now selects the normative clause -- the sentence carrying the sole
\MUST{} -- and runs all four assertions on that slice alone. Both halves of the
start rule are load-bearing and the helper's doc comment says why: last ". "
before the occurrence, falling back to just after the label only when there is
none. An unconditional fallback would return one slice spanning label to recap,
containing every needle; a period-only rule would swallow the label, which is
not a sentence.
Six mutations, six matching radii, test 3 alone in every case. Each attacks a
different layer: M22-M24 vary what the selected clause contains, M25 varies
which sentence is selected, M26 varies the fallback boundary -- starting after
the label versus widening before it -- and M27 varies whether selection is
unambiguous and therefore permitted at all. M27 is the one that reaches the
exactly-one assertion, because an assertion guarding the precondition of
selection cannot be exercised by any mutation that presupposes selection
succeeded.
Rows 4, 6, 7 and 8 only. core_spec.tex is not touched: the requirement's text
was always correct and only the guard was weak. The rung's original 38/38
matrix is preserved unchanged and amendment 3's evidence is appended beside it,
not merged into it.
One prose defect in ratified section 8.4 remains OPEN and is deliberately not
repaired here: that entry says M27 varies which clause is chosen, which M25 does
and M27 does not, since a correct implementation rejects M27 before selecting
any clause. Section 8.6 is unaffected and states something true. The annex
carries the corrected characterization and records the correction as open;
frozen text stays frozen regardless of severity, so the repair belongs to its
own amendment, which is not part of this commit.
Baseline unchanged at 43 suites, 1586 passing: amendment 3 alters an existing
test's scope and adds none.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ps1szk2mSfgp4Cz21eVH9x
Item 10 said only that "every cross-cutting structure's references resolve to
extant objects", naming no individual class, while the checker enforced a
surface spanning cross-cutting structures, structural top-level references,
meter references at every level a MeterChange appears, an attachment-internal
reference, event-internal references and the tempo map's segment anchors. The
Rust doc comment named many of those and not all. The two were incomplete in
different places, so neither could be repaired from the other; both are now
written from one table derived by reading every emitted condition in the four
functions that raise the tag -- 41 tokens, each with its resolution target and a
symbolic anchor back to the control flow that enforces it.
Two of the ledger row's own claims did not survive scoping and were corrected at
ratification rather than carried: the G3a aside is ambiguous, not false, and the
two-sided repair stands on incompleteness rather than on a falsehood.
Guarded by exact (token, target) set equality in a new testkit suite, against an
oracle validated before use. Ordering and vocabulary are separate assertions
because an out-of-vocabulary term sorts perfectly well. Duplicates are checked on
the raw extraction, which set comparison cannot see. Item 10's opening sentence
is the slice anchor as a complete literal, required to occur exactly once, so
pin 3's retention of it is machine-observed rather than asserted. t12 is narrowed
and renamed, not deleted: cargo test -p epiphany-core must still fail when the
doc block is destroyed, and testkit is another crate.
Chapter 3 gains req:time:aleatoric-reference-locality -- an aleatoric region's
ordering and bounds references must name events of that same region, a locality
rule the checker always enforced and no requirement stated. Its three count
constants were measured at execution, never predicted: 214/285/285 -> 215/286/286.
38 mutations, 38 matching radii, every one against the full workspace with
--no-fail-fast and restored by hand write-back. M3 is the single passing control:
with equality weakened to actual.is_subset(&expected), M1-B stops failing, which
is what makes exactness load-bearing rather than assumed. Two harness faults
halted the run and are recorded in the annex rather than smoothed over; the
second exposed a real weakness in the requirement guard, reported and left for
amendment 3.
Files two candidates this rung does not repair: P13-S29, the invariant-10 tag
multiplexing Chapter 3/4 failures through a public API and its Display; and
P13-S30, the repository-wide assumption that TeX is spelled exactly, whose
requirement-block branch is demonstrated by this rung's own M20.
Baseline 42 suites/1583 -> 43/1586: one new suite, three tests, none removed.
Clippy and the pinned fmt gate clean on 1.95.0; core_spec.pdf rebuilt with zero
undefined references.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ps1szk2mSfgp4Cz21eVH9x
Levi Neuwirth