Files
aevolution/docs/adr/0149-atomic-cellular-migration-publication.md
dingo 0f071f3f8b
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 2m21s
Foundation / checks (network) (push) Successful in 20m51s
Foundation / checks (client) (push) Successful in 22m2s
Foundation / checks (tsan) (push) Successful in 22m59s
Foundation / checks (postgres) (push) Successful in 53m40s
Foundation / checks (checks) (push) Successful in 1h7m19s
Foundation / postgres-tools (push) Successful in 29s
Foundation / server-image (push) Successful in 38s
Foundation / image (push) Successful in 4m31s
M12.4: atomically publish and recover biological migration pairs
2026-10-18 10:02:55 -04:00

3.5 KiB

ADR-0149: Atomic paired biological migration publication

Status

Accepted and locally qualified for scoped M12.4j implementation, 2026-10-18. Executed evidence and limits are in VALIDATION. No automatic routing, ecological boundary or deployment is selected.

Decision

Publish ADR-0144's two exact candidates in one PostgreSQL transaction after ADR-0147 preparation and ADR-0148 destination staging. Lock endpoint lease rows in increasing region order, require both current authorities and unchanged captured heads, and bind the destination payload to the immutable staged ghost. The immutable shared receipt names the preparation and both result commits. Both result manifests preserve their profiles, bundle 10, history, quantum and completed tick while advancing each commit sequence exactly once. They use no ordinary biological batch or invented birth/death event.

Restore both verified predecessor capsules, replay the exact trace and upload/ read back both complete results before publication. Clear both pending pointers, insert the common receipt, insert both commits and update both heads in the same transaction. Any failure rolls everything back, including a failure at the second commit. Deferred receipt foreign keys allow evidence guards to recognize this dedicated publisher without disabling triggers or permitting generic bypasses.

Before this commit the source owns all outgoing stock and the ghost owns none. After it only the destination owns the transferred organisms/material. Current regional budgets and explicit cumulative boundary flows retain their meanings. Keep both predecessor histories/capsules, stage and receipt. No stock account is created for the receipt, and no terminal extinction inference follows from it.

Explicit source abort and paired publication serialize on the source lease row. Abort wins only before commit; a committed request can never be aborted. If abort wins, commit is refused. Exact committed retries compare both complete manifests and return the original pair under current endpoint authority, including newer epochs; changed bytes conflict. Unknown replies retain that exact prepared pair. Lease replacement never changes its bytes or fabricates a second transfer.

Recovery loads the shared receipt under current local authority, verifies both predecessor/result manifests and their links, replays both candidates from local retained capsules, compares both result digests and the complete local payload, and requires the staged destination digest/length. Missing or false evidence fails closed. A paused host whose pointers disappeared reloads its actual head and performs this recovery before readiness; it never resumes stale source stock or activates a ghost based solely on pointer absence.

Alternatives, risks and qualification

Two ordinary publications can leave half a transfer committed. Treating timeout as rollback can duplicate stock. Shared append-only evidence plus one fenced transaction gives one outcome, at the cost of temporarily requiring both owners and one metadata authority. This is not a general distributed transaction across independent databases.

Require both representations, complete stock/member equality, exact/conflicting retries, abort/commit races, stale/replacement authorities, failed second inserts, ambiguous replies, process loss before/after commit, missing/corrupt/false replay evidence, host adoption/restart and subsequent biological continuation. Linked geometry, independent network partitions, actual multi-day ecology and target host capacity remain separate gates.