Incremental persistence: protocol obligations

Status: design constraints, not an implemented or validated storage backend. This accompanies the execution plan. The current reference remains synchronous whole-model localStorage persistence.

Observable reference boundary

An update executes once. spaTransition keeps the previous model if update or render fails before render commits. A post/Cmd failure after that commit keeps the next model. Persistence then attempts to write that kept model, and session-clear handling follows the existing encoder/write order. Persistence is best effort when the browser denies storage; this does not justify introducing additional loss windows while storage works.

Every step republishes its model, including an ignored key. In two tabs, A can overwrite B's last snapshot even when A's own model is unchanged. Avoiding an encode must therefore preserve publication; suppressing all no-op writes would change the existing last-writer semantics. Metadata, identity fencing and session-clear decisions remain separate obligations.

Why an asynchronous replacement is insufficient

Publishing an IndexedDB transaction after rendering introduces a visible edit that immediate reload can lose. Publishing before rendering can instead restore a model whose render failed, whereas reference execution retains the previous model. There is no transaction spanning DOM commit and IndexedDB commit.

Immutable localStorage chunks plus one manifest avoid neither cross-tab garbage collection races nor quota limits by themselves. An unpublished writer and a reader restoring an older root must be protected from collection. Wall-clock leases and arbitrarily long retention are not correctness proofs.

Candidate hybrid boundary

Prepare immutable chunks in IndexedDB while holding the existing message scheduler before model publication. Wait for transaction completion. Render and post using the existing panic/commit rules, then synchronously publish a small localStorage root for the model actually kept. The root is the publication boundary, not the preparatory database transaction. A no-op still republishes its root. Failed preparation must use the full synchronous reference path where that path remains available.

This candidate requires a coordinator shared by all readers, preparers, publishers and collectors in the same storage partition. It must protect roots until publication and pin boot readers until their read transaction finishes. An old tab must validate or recreate its retained root before republishing it after another tab has collected unreachable chunks. Identity changes cannot make another identity's data eligible for restore. Reference-format migration, mixed old/new tabs and unsupported browser capabilities need explicit rules.

Pre-protocol tab compatibility

An older protocol-aware tab can participate in coordination and revalidate its retained root. A pre-protocol tab cannot: it still reads and publishes full JSON at the existing model key and does not acquire the new locks.

A separate root key cannot preserve publication order with those tabs. Suppose the existing key contains model A, a new tab publishes root B without updating that key, and a suspended pre-protocol tab resumes and republishes A. The existing key still contains A; the root still names B. Republishing an identical value produces no storage event, so this state cannot reveal whether A was published before or after B. The HTML setItem algorithm returns before broadcasting when the stored value already equals the supplied value. Replacing the existing key with a root instead breaks pre-protocol restore.

The conservative compatibility policy is therefore to keep full JSON at the existing key as the sole publication authority while pre-protocol contexts may exist. New tabs must continue synchronous full-JSON publication, including unchanged steps, with the existing identity, sign-out and size-cap rules. An optional chunk sidecar may only cache the current authoritative value after content, schema and identity validation; it must not resurrect state after that value changes or disappears, or introduce additional storage-failure windows. This policy does not remove the cost of whole-model publication.

Root-only publication requires a storage scope known never to have hosted pre-protocol code, or an explicit compatibility cutover that accepts changed old-tab restore and last-writer semantics. Timeouts, heartbeats and a staged deployment do not prove that suspended pre-protocol writers are gone. These are design constraints and a candidate policy, not an implemented migration.

Web Locks associate coordination with a storage bucket and release locks on agent termination; they do not turn the two storage APIs into one transaction. Unsupported coordination must select reference execution rather than an unsafe chunk backend. Web Locks specification.

IndexedDB's complete event and transaction durability option have distinct meaning. The strict hint requests completion after persistent-medium writes; browser eviction and unavailable storage still require handling. Neither the hint nor a successful prepare publishes the application snapshot. IndexedDB specification.

Required falsification matrix before enabling

The hybrid remains a candidate until these obligations have executable proofs and browser evidence. No change in durability or multi-tab semantics is assumed to be authorised by the performance goal.