This is an infrastructure/plumbing change: it consolidates watch-triage, promotion, and agenda proposal/score decisions into a validated, hash-addressed, append-only store under proposals/intake/ and retires the legacy stores. No content/ records, claims, evidence, sources, or ratified assessments are touched, and the citation tooling found no added identifiers. Provenance is preserved rather than removed: legacy entries carry their original committed commit hash, ref, and verbatim record; migrate-intake.mjs verifies field-by-field replay and refuses to delete any legacy file that changed after the recorded basis (tested in intakeStore.test.ts), and the deleted YAMLs remain in git history. The checks are strengthened, not weakened: classify-pr-risk.mjs now forces needs-approval on any modification/deletion under proposals/intake/, incomplete (<4 seats) or stale panels are marked retryable and cannot yield an actionable score, and the four-highs/zero-concerns rule plus the separate publication gate are explicitly retained; the removed refusal-fallback ban comment is preserved in substance (seat identity kept, failures recorded as failed seats). Dropping the 'previously proposed title is permanently retired' heuristic is an editorial-workflow choice, not a constitutional check, and the replacement rests identical substance on identical case inputs while carrying prior reasons forward. Docs and the proposal page relabel 'retired' as 'not advanced' and surface seat reasoning honestly, noting that missing historical reasons stay unknown rather than being reconstructed. AGENTS.md is untouched; the DECISIONS.md addition is an AI implementation record in the established form.

The change replaces legacy intake stores with hash-validated, immutable batches while preserving each migrated row, its commit and locator, and leaving unavailable historical metadata null rather than reconstructing it. The migration verifies field-for-field replay before deleting old files and refuses deletion when a legacy file differs from the committed basis, satisfying provenance and history-preservation requirements. New panel records retain seat-level scores, reasoning, and concerns; score totals are validated against those seats, while stale or insufficient panels are marked retryable and excluded from actionable agenda tallies. Existing intake batches are explicitly removed from the low-risk lane when modified or deleted, and proposal-derived content still passes the independent publication gate rather than publishing on the generator's unilateral judgment.

The PR migrates scattered and ephemeral intake/agenda records into durable, schema-validated, and immutable batches under `proposals/intake/` while retiring legacy stores verified via replay. No content under `content/cases` is touched, no citations are added or fabricated, and the publication gates and 4-of-5 panel advancement requirements remain strictly enforced. Risk classification in `scripts/classify-pr-risk.mjs` is tightened rather than weakened by ensuring that any deletion or modification of existing intake history requires human approval. All domain tests pass and the changes preserve full inspectability and provenance across worker runs.

The diff does not edit AGENTS.md, add content/ claims or assessments, or introduce citations; citation tooling found none. It adds an append-only, hash-validated intake store under proposals/intake/, migrates legacy watch/promotion/agenda decisions with committed locators and unknown historical metadata left null, and only then retires the old ledgers—preserving provenance rather than fabricating or silently rewriting judgments (§3.8, §3.14, §3.15). classify-pr-risk now forces consequential review for any modify/delete of existing intake batches, incomplete or stale panels stay non-actionable, seat totals must match recorded seats, and the four-highs/zero-concerns rule plus the independent publication gate are unchanged. UI/copy shifts (“retired” → “not advanced,” optional review disclosure) and allowing same-title reconsideration only when substance or case inputs change do not weaken constitutional checks or present unratified drafts as ratified. No confidence material is published; Deep Memory remains explicitly unimported.

This is a workflow-infrastructure change with no new citations, claims, or assessments: the citation verification found no added identifiers, and the diff touches no content/ research records. The legacy decision stores are deleted only after a verified replay that preserves original rows with committed locators (legacy.commit/legacy.ref/legacy.record), keeps missing historical reviewer reasons as null rather than fabricating them, and refuses to retire any legacy file changed after its basis — consistent with §3.8/§3.9 and tested explicitly. The four-highs/zero-concerns advancement rule, the independent publication gate, and the consequential-change classification are retained or strengthened (modifications/deletions of existing intake batches now trigger needs-approval, per classify-pr-risk.mjs and tests), so no check is weakened or dodged; the 'not advanced' relabel is a more honest status than 'retired by silence', and reconsideration is gated on changed substance or inputs, not on renaming. No AGENTS.md edit appears in the diff. Residual uncertainties (whether the 195-record migration batches actually match the committed snapshot, PRs #175–#177 review claims) are migration-execution facts I cannot verify from the diff, but the replay-verification and idempotency logic shown is exactly the fail-closed pattern the constitution requires, and no visible record is fabricated or silently altered.