This is a pipeline/code change with no content records, no citations added (verification found none), and no edit to AGENTS.md. The substantive changes strengthen rather than dodge the checks: a settlement now snapshots the case's ledger YAML and restores it on any failure, the case must load as written or the run fails, a record already settled this pass is not re-written by the cascade, and a failed re-reading suppresses the edition re-telling instead of narrating over a half-written ledger. The `splitNoteFor` fix removes exactly the kind of fabricated provenance §3.8 forbids — a refusal note naming a part id (VASO-E087) that was never appended. The two judgment calls I weighed: (a) holding a live claim found compound when no part was admitted leaves a non-atomic claim standing (§3.2), but it is not hidden — it is written to the history line ("Left as they stood: …"), the run summary count (`held N`), and a "## Left as they stood" section of verification.md; (b) relinking a citing record to all parts could let evidence for one rung count for another (§3.2), but the appended limitation states explicitly that which part it bears on is for a later reading, so it is disclosed, not silent. Collapsing all whitespace in `normalizeForMatch` loosens quote matching slightly, but the motivating case (XML "p= 0.006" vs HTML "p = 0.006") is a rendering artifact, and the added test confirms a substantive wording difference still fails. Nothing here weakens §3.15's ratification or parking rules.

The new `plan.refuse.filter` deliberately leaves a claim live after the reader has found it compound whenever no split part is admitted; recording it as “held” does not cure the §3.2 violation that a Claim must contain one proposition. The `relive` cascade also attaches an existing Evidence record to every split part while explicitly saying that which parts it bears on remains for later reading, so evidential links and their inherited directions are asserted before they are verified under §3.5. In addition, `normalizeForMatch` removes all whitespace rather than normalizing whitespace boundaries, allowing distinct text such as “now here” and “nowhere” to pass the quotation check. These are substantive loosenings of atomicity, evidence-link, and quotation checks prohibited by §3.15. The change should retain failed splits only in a non-weight-bearing unresolved/provisional state, relink evidence only after the applicable parts are established, and normalize spacing without erasing token boundaries.

The change improves pipeline robustness during record re-reading and settlement without altering AGENTS.md or compromising epistemic checks. It ensures that compound claims without admitted parts are held rather than stripped away to cascade failures, properly relinks dependent records with explicit provenance notes, sanitizes split notes to reference only existing records, and adds transactional rollback if ledger writes fail validation. Normalizing whitespace in normalizeForMatch strictly aids matching between different electronic formats (such as XML and HTML) without altering recorded quotations.

The diff is pipeline/test/decision-log only: no AGENTS.md edit, no content citations, and no confidence material. It hardens answer-time re-reading so a live compound claim is held (not rejected into a hole) when no parts admit, citers relink to admitted parts with an explicit later-reading limitation, split notes name only appended ids, failed settlements snapshot-restore ledger YAML, and quote match ignores whitespace—each covered by tests and stamped in history/verification.md. The DECISIONS.md entry records a founder session direction, which §3.15 allows. None of this fabricates provenance, weakens §3.15 checks, or presents unratified judgments as ratified.

This is a pipeline-reliability change, not a content addition: citation verification finds no identifiers added, and no AGENTS.md edit occurs, so §3.15's constitution rule and the fabrication rules are not engaged. The four behaviors are epistemically honest: a compound-claim finding with no admitted parts is left displayed in the history, verification.md, and summary ("held"), rather than silently dropped or silently kept — the atomicity finding under §3.2 is recorded, not hidden; relinking a record to all parts of a split claim carries an explicit limitation that which part it bears on is undetermined, so evidence does not *silently* count for each rung, satisfying §3.2's proviso; and split notes naming only appended parts prevent pointing readers at records that do not exist, which serves §3.8's no-invented-locator rule rather than weakening it. The whitespace-insensitive quote matching is a legitimate normalization for rendering artifacts (the check still rejects genuinely altered quotes, per the third test), not a dodge of verbatim-quote verification. Rollback-and-report on failure, plus refusing to re-tell an edition over a rolled-back ledger, strengthens fail-closed integrity and §3.14's append-only history, with provenance (runId, date, actor, aiAssisted) preserved in every history entry.