This is pipeline/infrastructure work only: no content/ records are added or changed, and the citation tooling found no identifiers in added content lines. AGENTS.md is untouched; the protocol version bumps (draft-v7→v8, edition-v5→v6) archive the old files rather than deleting them, preserving the record. The substantive changes all tighten rather than loosen §3 checks: a fail-closed loader check (danglingProseRefs) that rejects prose naming non-existent record ids, a draft-verb decline for statements bundling a falsification clause (§3.2), narrowing a split evidence part's locator to the page its own quote sits on (§3.8), and a scheduler exclusion for cases with open sittings. The one arguably permissive element — PROSE_REFS_REQUIRED_FROM = 2026-09-12, exempting the 39 pre-existing dangling references — is disclosed in DECISIONS.md with the affected cases and left for a correction pass, which is a new check scoped forward, not an existing check weakened. The exactLocator reader act is stamped honestly (model field explicitly reads "verify (mechanical: the page marker before the part's quote)"), so no human or model judgment is misattributed. The PR body's claims of founder direction are recorded as an operator record with date and channel, consistent with §3.15's operating-record clause.

The new dangling-prose-reference gate is deliberately not fail-closed across the repository: `checkIntegrity` passes `PROSE_REFS_REQUIRED_FROM`, exempting every record dated before 2026-09-12. `docs/DECISIONS.md` acknowledges 39 existing dangling references, while the constitution requires CI to enforce dangling-ID validity and separately says “No dangling IDs.” Because eligibility depends on an editable provenance date, a newly introduced or modified record can also evade the check by carrying an earlier date, making this a dodge path rather than merely deferred cleanup. Remove the cutoff or first repair/explicitly supersede the known dangling references, and make the check apply to all relevant records regardless of origin date.

The change introduces upstream mechanical validation rules learned from previous panel reviews, directly reinforcing constitutional requirements. Specifically, it enforces §3.2 by mechanically rejecting claim statements that bundle falsification conditions into compound propositions at the draft phase, enforces §3.8 by narrowing exact locators to the specific page of split evidence quotes (with recorded provenance under readerActs), and prevents dangling prose references to nonexistent records in the loader and drafter. AGENTS.md is untouched, no citations or records are fabricated, and the pipeline checks are strengthened rather than weakened.

The diff is infrastructure, protocols, tests, and DECISIONS/AUTOMATION only — no content/ ledger records, no AGENTS.md edit, no citations added, and no run account to judge. It hardens existing §3 rules fail-closed: draft declines compound falsification clauses (§3.2) and prose that names non-ledger ids; verify narrows split-part exactLocator to the quote’s page and stamps a readerAct on the new schema field (§3.8); load fails dangling prose refs from 2026-09-12 with the pre-cutoff backlog listed honestly in DECISIONS rather than hidden. Busy-case exclusion and edition word-count notes are operational, not check-weakening. Founder direction is recorded as an operator decision stamp, which §3.15 allows.

This is a tooling and protocol change, not a content publication: no citations are added (the verification section confirms none), AGENTS.md is untouched, and no unratified draft is presented as ratified. The five fixes strengthen rather than weaken §3.15's checks — dangling prose references now fail the build, compound §3.2 statements are declined at draft, locators are narrowed mechanically with the change stamped as a reader act (from/to/model/runId recorded, §3.14 provenance preserved), old protocols are archived rather than deleted, and the scheduler change fails safe (with no `gh` token, every unmerged recent branch counts as busy). The prospective cutoff (PROSE_REFS_REQUIRED_FROM = 2026-09-12) leaves a 39-item backlog unchecked rather than fixing it, which deserves an operator answer — but it adds a new check with a documented transition and a recorded correction plan in DECISIONS.md; it does not loosen or reclassify any check the constitution already mandates, so it is a transition choice, not check-weakening. The claimed founder direction is recorded in DECISIONS.md with date and channel, matching the constitution's operating-record pattern, and the falsification-clause rejection is a defensible reading of §3.2 applied with recorded dispositions, not silent drops.