[Voice] Preserve imported primary Voice change history at knowledge cutoffs

[seonghobae]

Product gap

ADR 0251 prevents a later imported primary Voice from being antedated, but migration 0237 currently deletes the former primary assignment before inserting the new one. An authorized historical-cutoff read therefore cannot recover the former primary after a source update. It honestly returns unavailable, but it does not preserve the product history users expect.

Required decision

Define an ADR-first temporal persistence contract before changing the schema. Reuse the existing source_post_voice boundary and PostgreSQL range/constraint features where possible; do not infer a Voice, invent a weight, or duplicate the source representation. Keep additional evidence-bearing assignments and counterparty relationships separate.

Acceptance

• A primary Voice update closes the former effective interval and opens the new interval atomically.
• Live reads return exactly one current imported primary; an authorized cutoff read returns the primary effective at that instant.
• Re-import and migration replay are idempotent and do not manufacture history that the source never supplied.
• The schema permits repeated use of the same atomic Voice in non-overlapping periods while rejecting overlapping primary intervals.
• API, ontology JSON-LD/CSV, and UI preserve ABAC and cutoff boundaries; hidden evidence remains hidden.
• Synthetic PostgreSQL integration tests prove A → B → A changes at before/between/after cutoffs and concurrent UPSERT behavior.
• ADR 0251, PRD/TRD/UML, and docs/product-technical-gap-baseline.md distinguish non-antedating from historical preservation.

No real organization or record identifiers belong in fixtures or evidence.

[seonghobae]

Cycle 2026-08-28: GitHub writes recovered (comments 201, close-without-reason 200, admin on REST+GraphQL). Token still has empty X-OAuth-Scopes; git clone via helper still 401, public clone used. Independent APPROVE remains required — this comment does not approve or merge.

Open PRs 702/679/672/668/667/640 are all dirty vs current main (bbb19192, v2.22.0). None has an independent APPROVE on the exact head. Copilot review requested on #702. #96 closed as the weaker duplicate of merged #91.

Next buyer gap owned here: implement ADR 0252 temporal primary Voice history (close former primary interval, open the new one atomically; cutoff reads recover the former primary). Distinct from ADR 0251 non-antedating. Synthetic fixtures only.

[seonghobae]

Closing as completed: ADR 0252 is Accepted on protected main and explicitly closes this issue. Temporal primary Voice history landed via #761 (half-open intervals, cutoff reads, A→B→A). No further unique slice invented from this issue.

[seonghobae]

Live synthetic PostgreSQL proof for ADR 0252 is in #763 (v2.22.1, feat/voice-history-live-postgres-v2221 @ 1c9986c).

The tests cover A → B → A at before/between/after cutoffs, concurrent voc_type_code updates, GiST non-overlap, closing a matching additional assignment, and 0237→0243 trigger replay. They skip without PostgreSQL.

Do not close this issue until #763 lands through the protected squash-merge gate with independent exact-head APPROVE. This session will not self-approve.