The change repairs prose that named record ids the ledger does not hold, and removes the date cutoff from `danglingProseRefs` so the no-dangling-id rule (§8) applies to every record — this tightens rather than weakens a check (§3.15), and the remaining optional `since` parameter is test-only tooling. No AGENTS.md edit; no new citations except the pre-existing Janusek DOI, which the tooling resolves to 'Building Taypikala' as cited. Every content edit is mirrored by an append-only history entry giving the exact old → new text, a per-record reason, the actor (claude-fable-5-1, operator) and aiAssisted:true, satisfying §3.14; nothing is presented as a ratified or human judgment. The verify-pipeline fix (scrubUnadmitted) addresses the generating cause and writes a run note when it rewrites, which is honest provenance. My one reservation, recorded as a note rather than an objection: the edits rewrite text attributed in-line to a named second reader ('claude-sonnet-5, 2026-09-10, run …'), so those quoted notes no longer read exactly as that reader wrote them (e.g. CCC-E057, TRN-E079); the substitutions are semantically faithful and the originals are preserved verbatim in history, but I would prefer the altered passages carry an in-place marker (e.g. bracketed editorial substitution) so a reader sees the reader's words were edited without consulting the changelog. That is a refinement, not a rule break.

The change removes the publication loader’s date cutoff and makes dangling prose references fail across all records, strengthening rather than weakening the fail-closed integrity check. Existing invalid references are either redirected to identified ledger records or replaced with prose explicitly saying that the referenced proposal was not admitted, so rejected draft records are not presented as standing records. The content corrections are accompanied by append-only history entries recording before-and-after text, reasons, date, model/operator attribution, and preserved reader disagreement. The only mechanically checked citation resolves to the matching “Building Taypikala” chapter, and the diff adds no AGENTS.md edit, confidence material, fabricated identifier, or silent ratification.

The change directly addresses review note #285 by removing the artificial date cutoff on dangling prose ID checks, enforcing §8's 'no dangling IDs' universally across all cases. It meticulously repairs legacy dangling references in evidence and research prose across six cases (ccc, deep-memory, geopolymer, pre-columbian-amazon, transients, vasocomputation), re-pointing to existing records where appropriate or describing unadmitted proposal IDs accurately. Each edit is properly recorded in the respective case's history.yaml, shorthand ID references are now caught by the validator, and verify.ts is enhanced with an automated scrub to prevent future intake notes from referencing unadmitted proposal IDs.

The change answers review note #285 by removing the prose-ref date cutoff so danglingProseRefs runs on every record, adding same-case shorthand detection, and repairing the named legacy prose refs with append-only history entries that stamp actor, date, and from→to reasons. That strengthens §8 / §3.15 mechanical checks rather than weakening them; AGENTS.md is untouched. Re-points go to ledger ids the PR asserts exist; unadmitted proposal ids are rewritten as explicit non-admitted prose, and verify gains scrubUnadmitted so the pipeline cannot re-seed dangling proposal ids in second-reader notes. The sole added/touched DOI resolves; no confidence material or fabricated citations appear in the diff. Ordinary awkward residual wording is not a constitutional break.

The change strengthens, not weakens, the §8 'No dangling IDs' rule: it removes a date cutoff that was a dodge path, extends the loader's prose-reference check to every record, and adds a fail-closed CI pass (build reports zero dangling references, and the loader check is throw-on-dangling). The 35 legacy references were repaired through the correction writer with append-only history entries recording each from → to, the reason, actor, date, and aiAssisted stamp, satisfying §3.14's versioning requirement; re-pointings (e.g. GEO-E706 → GEO-E002, CCC-E041 → CCC-E067) are justified in the history text by what the ledger record holds, and rewrites replace ids of records that never entered with honest descriptive phrases ('a proposed claim not admitted at intake') rather than inventing records or locators. The verify.ts scrub is mechanical, grounded in the run's own proposal files (the quoted 'observed' text comes from the actual proposal records), and logged in run notes, so it does not misattribute the second reader's findings — it makes them nameable without dangling ids. AGENTS.md is untouched; docs/DECISIONS.md is the decisions log, which the constitution expressly uses. The one DOI appearing in changed prose (Janusek et al. 2013) mechanically resolves and matches its described content. Minor residual items — e.g. DEEP-E023's origin.ref still naming the provisional parent DEEP-E009 — are historical lineage records of a run, not prose claims that a ledger record exists, and at most warrant a review note, not a violation.