perf(client): batch large thread sync updates (upstream #5344)

[patroza]

Imports open upstream PR pingdotgg/t3code#5344 as one provenance commit on fork/candidates.

Source SHA: 783fd023c69f7ebc9c0ae657ce124cce77d94587 (upstream commits b623dc23b + 783fd023c, squashed).

Imported behavior

• reduceThreadStreamItems: a pure reducer that folds a batch of OrchestrationThreadStreamItems into one state and one persistable snapshot, preserving sequence dedup, completion markers, and deletion handling.
• Stream.groupedWithin(64, 16ms) on the live thread subscription, so a burst of thread events publishes the SubscriptionRefonce instead of once per event. Web and mobile stop rebuilding large thread views for every streamed event.
• eventBatchSize on EnvironmentThreadStateOptions, plus upstream's two regression tests: ordered single-publication bursts, and persisting a settled snapshot when a batch ends with a non-persistable turn start.

Local adaptations

• Kept our httpSnapshotLoadAttempted guard around the HTTP snapshot fallback (upstream does not have it); the call now goes through applyItems([...]).
• Restored setDeleted, which upstream deleted. Our terminal thread-deleted subscription failure (terminalSnapshotReason) never reaches the item stream, so it cannot go through the batch reducer. Cache removal is now shared between that path and the reducer path via removeCachedThread.

Excluded

• tasks/todo.md — the upstream author's scratch checklist.

Follow-up (belongs to fork/changes, not this layer)

fork/changes adds reloadFromServer and the result.kind === "reload-required" branch on top of setThread, which this import removes. Rebasing fork/changes onto this layer will conflict there (deliberately — it must not be resolved by dropping the branch). The intended re-expression: have the reducer report the reload-required index, commit state up to it, run reloadFromServer(), then re-enter applyItems with the remaining items — with reloadFromServer itself calling applyItems([{ kind: "snapshot", snapshot: fresh.value }]) instead of the deleted setThread.

Risks worth knowing

• Live state is now coalesced up to 16 ms (~1 frame). Under TestClock, a batch smaller than 64 items only flushes when the clock advances; upstream's harness pins eventBatchSize: 1 so existing tests keep their per-item assertions.
• applyItems snapshots state / awaitingCompletion at batch start and writes them back at the end, so a concurrent setDisconnected / setStreamError from the supervisor fiber can be clobbered. The race is pre-existing, but the window widens from per-event to ~16 ms.

Validation

• vp test run packages/client-runtime/src/state/{threads-sync,threadReducer,threads-atoms}.test.ts — 52 passed
• vp run -F @t3tools/client-runtime -F @t3tools/web test — 210 files / 1874 tests passed
• vp check — 0 errors
• vp run -F @t3tools/client-runtime -F @t3tools/contracts -F @t3tools/web -F @t3tools/mobile typecheck — clean

Workspace-wide typecheck is not meaningful locally on this layer: this worktree's node_modules was installed from fork/changes (@anthropic-ai/claude-agent-sdk@^0.3.220) while fork/candidates pins ^0.3.170, which produces unrelated ClaudeAdapter type errors in apps/server. Fork CI installs from this branch's own lockfile, so it is the authority there. This change touches only packages/client-runtime/src/state/ and the candidate registry.

Note: the agent ship-gate tooling (.githooks/, scripts/agent-pre-push.mjs, pnpm pr:ready) lives in fork/changes and does not exist on fork/candidates, so the equivalent checks above were run by hand.

[patroza]

Imports open upstream PR pingdotgg/t3code#5344 as one provenance commit on fork/candidates.

Source SHA: 783fd023c69f7ebc9c0ae657ce124cce77d94587 (upstream commits b623dc23b + 783fd023c, squashed).

Imported behavior

• reduceThreadStreamItems: a pure reducer that folds a batch of OrchestrationThreadStreamItems into one state and one persistable snapshot, preserving sequence dedup, completion markers, and deletion handling.
• Stream.groupedWithin(64, 16ms) on the live thread subscription, so a burst of thread events publishes the SubscriptionRefonce instead of once per event. Web and mobile stop rebuilding large thread views for every streamed event.
• eventBatchSize on EnvironmentThreadStateOptions, plus upstream's two regression tests: ordered single-publication bursts, and persisting a settled snapshot when a batch ends with a non-persistable turn start.

Local adaptations

• Kept our httpSnapshotLoadAttempted guard around the HTTP snapshot fallback (upstream does not have it); the call now goes through applyItems([...]).
• Restored setDeleted, which upstream deleted. Our terminal thread-deleted subscription failure (terminalSnapshotReason) never reaches the item stream, so it cannot go through the batch reducer. Cache removal is now shared between that path and the reducer path via removeCachedThread.

Excluded

• tasks/todo.md — the upstream author's scratch checklist.

Follow-up (belongs to fork/changes, not this layer)

fork/changes adds reloadFromServer and the result.kind === "reload-required" branch on top of setThread, which this import removes. Rebasing fork/changes onto this layer will conflict there (deliberately — it must not be resolved by dropping the branch). The intended re-expression: have the reducer report the reload-required index, commit state up to it, run reloadFromServer(), then re-enter applyItems with the remaining items — with reloadFromServer itself calling applyItems([{ kind: "snapshot", snapshot: fresh.value }]) instead of the deleted setThread.

Risks worth knowing

• Live state is now coalesced up to 16 ms (~1 frame). Under TestClock, a batch smaller than 64 items only flushes when the clock advances; upstream's harness pins eventBatchSize: 1 so existing tests keep their per-item assertions.
• applyItems snapshots state / awaitingCompletion at batch start and writes them back at the end, so a concurrent setDisconnected / setStreamError from the supervisor fiber can be clobbered. The race is pre-existing, but the window widens from per-event to ~16 ms.

Validation

• vp test run packages/client-runtime/src/state/{threads-sync,threadReducer,threads-atoms}.test.ts — 52 passed
• vp run -F @t3tools/client-runtime -F @t3tools/web test — 210 files / 1874 tests passed
• vp check — 0 errors
• vp run -F @t3tools/client-runtime -F @t3tools/contracts -F @t3tools/web -F @t3tools/mobile typecheck — clean

Workspace-wide typecheck is not meaningful locally on this layer: this worktree's node_modules was installed from fork/changes (@anthropic-ai/claude-agent-sdk@^0.3.220) while fork/candidates pins ^0.3.170, which produces unrelated ClaudeAdapter type errors in apps/server. Fork CI installs from this branch's own lockfile, so it is the authority there. This change touches only packages/client-runtime/src/state/ and the candidate registry.

Note: the agent ship-gate tooling (.githooks/, scripts/agent-pre-push.mjs, pnpm pr:ready) lives in fork/changes and does not exist on fork/candidates, so the equivalent checks above were run by hand.

[patroza]

Imports open upstream PR pingdotgg/t3code#5344 as one provenance commit on fork/candidates.

Source SHA: 783fd023c69f7ebc9c0ae657ce124cce77d94587 (upstream commits b623dc23b + 783fd023c, squashed).

Imported behavior

• reduceThreadStreamItems: a pure reducer that folds a batch of OrchestrationThreadStreamItems into one state and one persistable snapshot, preserving sequence dedup, completion markers, and deletion handling.
• Stream.groupedWithin(64, 16ms) on the live thread subscription, so a burst of thread events publishes the SubscriptionRefonce instead of once per event. Web and mobile stop rebuilding large thread views for every streamed event.
• eventBatchSize on EnvironmentThreadStateOptions, plus upstream's two regression tests: ordered single-publication bursts, and persisting a settled snapshot when a batch ends with a non-persistable turn start.

Local adaptations

• Kept our httpSnapshotLoadAttempted guard around the HTTP snapshot fallback (upstream does not have it); the call now goes through applyItems([...]).
• Restored setDeleted, which upstream deleted. Our terminal thread-deleted subscription failure (terminalSnapshotReason) never reaches the item stream, so it cannot go through the batch reducer. Cache removal is now shared between that path and the reducer path via removeCachedThread.

Excluded

• tasks/todo.md — the upstream author's scratch checklist.

Follow-up (belongs to fork/changes, not this layer)

fork/changes adds reloadFromServer and the result.kind === "reload-required" branch on top of setThread, which this import removes. Rebasing fork/changes onto this layer will conflict there (deliberately — it must not be resolved by dropping the branch). The intended re-expression: have the reducer report the reload-required index, commit state up to it, run reloadFromServer(), then re-enter applyItems with the remaining items — with reloadFromServer itself calling applyItems([{ kind: "snapshot", snapshot: fresh.value }]) instead of the deleted setThread.

Risks worth knowing

• Live state is now coalesced up to 16 ms (~1 frame). Under TestClock, a batch smaller than 64 items only flushes when the clock advances; upstream's harness pins eventBatchSize: 1 so existing tests keep their per-item assertions.
• applyItems snapshots state / awaitingCompletion at batch start and writes them back at the end, so a concurrent setDisconnected / setStreamError from the supervisor fiber can be clobbered. The race is pre-existing, but the window widens from per-event to ~16 ms.

Validation

• vp test run packages/client-runtime/src/state/{threads-sync,threadReducer,threads-atoms}.test.ts — 52 passed
• vp run -F @t3tools/client-runtime -F @t3tools/web test — 210 files / 1874 tests passed
• vp check — 0 errors
• vp run -F @t3tools/client-runtime -F @t3tools/contracts -F @t3tools/web -F @t3tools/mobile typecheck — clean

Workspace-wide typecheck is not meaningful locally on this layer: this worktree's node_modules was installed from fork/changes (@anthropic-ai/claude-agent-sdk@^0.3.220) while fork/candidates pins ^0.3.170, which produces unrelated ClaudeAdapter type errors in apps/server. Fork CI installs from this branch's own lockfile, so it is the authority there. This change touches only packages/client-runtime/src/state/ and the candidate registry.

Note: the agent ship-gate tooling (.githooks/, scripts/agent-pre-push.mjs, pnpm pr:ready) lives in fork/changes and does not exist on fork/candidates, so the equivalent checks above were run by hand.

[patroza]

Imports open upstream PR pingdotgg/t3code#5344 as one provenance commit on fork/candidates.

Source SHA: 783fd023c69f7ebc9c0ae657ce124cce77d94587 (upstream commits b623dc23b + 783fd023c, squashed).

Imported behavior

• reduceThreadStreamItems: a pure reducer that folds a batch of OrchestrationThreadStreamItems into one state and one persistable snapshot, preserving sequence dedup, completion markers, and deletion handling.
• Stream.groupedWithin(64, 16ms) on the live thread subscription, so a burst of thread events publishes the SubscriptionRefonce instead of once per event. Web and mobile stop rebuilding large thread views for every streamed event.
• eventBatchSize on EnvironmentThreadStateOptions, plus upstream's two regression tests: ordered single-publication bursts, and persisting a settled snapshot when a batch ends with a non-persistable turn start.

Local adaptations

• Kept our httpSnapshotLoadAttempted guard around the HTTP snapshot fallback (upstream does not have it); the call now goes through applyItems([...]).
• Restored setDeleted, which upstream deleted. Our terminal thread-deleted subscription failure (terminalSnapshotReason) never reaches the item stream, so it cannot go through the batch reducer. Cache removal is now shared between that path and the reducer path via removeCachedThread.

Excluded

• tasks/todo.md — the upstream author's scratch checklist.

Follow-up (belongs to fork/changes, not this layer)

fork/changes adds reloadFromServer and the result.kind === "reload-required" branch on top of setThread, which this import removes. Rebasing fork/changes onto this layer will conflict there (deliberately — it must not be resolved by dropping the branch). The intended re-expression: have the reducer report the reload-required index, commit state up to it, run reloadFromServer(), then re-enter applyItems with the remaining items — with reloadFromServer itself calling applyItems([{ kind: "snapshot", snapshot: fresh.value }]) instead of the deleted setThread.

Risks worth knowing

• Live state is now coalesced up to 16 ms (~1 frame). Under TestClock, a batch smaller than 64 items only flushes when the clock advances; upstream's harness pins eventBatchSize: 1 so existing tests keep their per-item assertions.
• applyItems snapshots state / awaitingCompletion at batch start and writes them back at the end, so a concurrent setDisconnected / setStreamError from the supervisor fiber can be clobbered. The race is pre-existing, but the window widens from per-event to ~16 ms.

Validation

• vp test run packages/client-runtime/src/state/{threads-sync,threadReducer,threads-atoms}.test.ts — 52 passed
• vp run -F @t3tools/client-runtime -F @t3tools/web test — 210 files / 1874 tests passed
• vp check — 0 errors
• vp run -F @t3tools/client-runtime -F @t3tools/contracts -F @t3tools/web -F @t3tools/mobile typecheck — clean

Workspace-wide typecheck is not meaningful locally on this layer: this worktree's node_modules was installed from fork/changes (@anthropic-ai/claude-agent-sdk@^0.3.220) while fork/candidates pins ^0.3.170, which produces unrelated ClaudeAdapter type errors in apps/server. Fork CI installs from this branch's own lockfile, so it is the authority there. This change touches only packages/client-runtime/src/state/ and the candidate registry.

Note: the agent ship-gate tooling (.githooks/, scripts/agent-pre-push.mjs, pnpm pr:ready) lives in fork/changes and does not exist on fork/candidates, so the equivalent checks above were run by hand.

[patroza]

Imports open upstream PR pingdotgg/t3code#5344 as one provenance commit on fork/candidates.

Source SHA: 783fd023c69f7ebc9c0ae657ce124cce77d94587 (upstream commits b623dc23b + 783fd023c, squashed).

Imported behavior

• reduceThreadStreamItems: a pure reducer that folds a batch of OrchestrationThreadStreamItems into one state and one persistable snapshot, preserving sequence dedup, completion markers, and deletion handling.
• Stream.groupedWithin(64, 16ms) on the live thread subscription, so a burst of thread events publishes the SubscriptionRefonce instead of once per event. Web and mobile stop rebuilding large thread views for every streamed event.
• eventBatchSize on EnvironmentThreadStateOptions, plus upstream's two regression tests: ordered single-publication bursts, and persisting a settled snapshot when a batch ends with a non-persistable turn start.

Local adaptations

• Kept our httpSnapshotLoadAttempted guard around the HTTP snapshot fallback (upstream does not have it); the call now goes through applyItems([...]).
• Restored setDeleted, which upstream deleted. Our terminal thread-deleted subscription failure (terminalSnapshotReason) never reaches the item stream, so it cannot go through the batch reducer. Cache removal is now shared between that path and the reducer path via removeCachedThread.

Excluded

• tasks/todo.md — the upstream author's scratch checklist.

Follow-up (belongs to fork/changes, not this layer)

fork/changes adds reloadFromServer and the result.kind === "reload-required" branch on top of setThread, which this import removes. Rebasing fork/changes onto this layer will conflict there (deliberately — it must not be resolved by dropping the branch). The intended re-expression: have the reducer report the reload-required index, commit state up to it, run reloadFromServer(), then re-enter applyItems with the remaining items — with reloadFromServer itself calling applyItems([{ kind: "snapshot", snapshot: fresh.value }]) instead of the deleted setThread.

Risks worth knowing

• Live state is now coalesced up to 16 ms (~1 frame). Under TestClock, a batch smaller than 64 items only flushes when the clock advances; upstream's harness pins eventBatchSize: 1 so existing tests keep their per-item assertions.
• applyItems snapshots state / awaitingCompletion at batch start and writes them back at the end, so a concurrent setDisconnected / setStreamError from the supervisor fiber can be clobbered. The race is pre-existing, but the window widens from per-event to ~16 ms.

Validation

• vp test run packages/client-runtime/src/state/{threads-sync,threadReducer,threads-atoms}.test.ts — 52 passed
• vp run -F @t3tools/client-runtime -F @t3tools/web test — 210 files / 1874 tests passed
• vp check — 0 errors
• vp run -F @t3tools/client-runtime -F @t3tools/contracts -F @t3tools/web -F @t3tools/mobile typecheck — clean

Workspace-wide typecheck is not meaningful locally on this layer: this worktree's node_modules was installed from fork/changes (@anthropic-ai/claude-agent-sdk@^0.3.220) while fork/candidates pins ^0.3.170, which produces unrelated ClaudeAdapter type errors in apps/server. Fork CI installs from this branch's own lockfile, so it is the authority there. This change touches only packages/client-runtime/src/state/ and the candidate registry.

Note: the agent ship-gate tooling (.githooks/, scripts/agent-pre-push.mjs, pnpm pr:ready) lives in fork/changes and does not exist on fork/candidates, so the equivalent checks above were run by hand.

[patroza]

Imports open upstream PR pingdotgg/t3code#5344 as one provenance commit on fork/candidates.

Source SHA: 783fd023c69f7ebc9c0ae657ce124cce77d94587 (upstream commits b623dc23b + 783fd023c, squashed).

Imported behavior

• reduceThreadStreamItems: a pure reducer that folds a batch of OrchestrationThreadStreamItems into one state and one persistable snapshot, preserving sequence dedup, completion markers, and deletion handling.
• Stream.groupedWithin(64, 16ms) on the live thread subscription, so a burst of thread events publishes the SubscriptionRefonce instead of once per event. Web and mobile stop rebuilding large thread views for every streamed event.
• eventBatchSize on EnvironmentThreadStateOptions, plus upstream's two regression tests: ordered single-publication bursts, and persisting a settled snapshot when a batch ends with a non-persistable turn start.

Local adaptations

• Kept our httpSnapshotLoadAttempted guard around the HTTP snapshot fallback (upstream does not have it); the call now goes through applyItems([...]).
• Restored setDeleted, which upstream deleted. Our terminal thread-deleted subscription failure (terminalSnapshotReason) never reaches the item stream, so it cannot go through the batch reducer. Cache removal is now shared between that path and the reducer path via removeCachedThread.

Excluded

• tasks/todo.md — the upstream author's scratch checklist.

Follow-up (belongs to fork/changes, not this layer)

fork/changes adds reloadFromServer and the result.kind === "reload-required" branch on top of setThread, which this import removes. Rebasing fork/changes onto this layer will conflict there (deliberately — it must not be resolved by dropping the branch). The intended re-expression: have the reducer report the reload-required index, commit state up to it, run reloadFromServer(), then re-enter applyItems with the remaining items — with reloadFromServer itself calling applyItems([{ kind: "snapshot", snapshot: fresh.value }]) instead of the deleted setThread.

Risks worth knowing

• Live state is now coalesced up to 16 ms (~1 frame). Under TestClock, a batch smaller than 64 items only flushes when the clock advances; upstream's harness pins eventBatchSize: 1 so existing tests keep their per-item assertions.
• applyItems snapshots state / awaitingCompletion at batch start and writes them back at the end, so a concurrent setDisconnected / setStreamError from the supervisor fiber can be clobbered. The race is pre-existing, but the window widens from per-event to ~16 ms.

Validation

• vp test run packages/client-runtime/src/state/{threads-sync,threadReducer,threads-atoms}.test.ts — 52 passed
• vp run -F @t3tools/client-runtime -F @t3tools/web test — 210 files / 1874 tests passed
• vp check — 0 errors
• vp run -F @t3tools/client-runtime -F @t3tools/contracts -F @t3tools/web -F @t3tools/mobile typecheck — clean

Workspace-wide typecheck is not meaningful locally on this layer: this worktree's node_modules was installed from fork/changes (@anthropic-ai/claude-agent-sdk@^0.3.220) while fork/candidates pins ^0.3.170, which produces unrelated ClaudeAdapter type errors in apps/server. Fork CI installs from this branch's own lockfile, so it is the authority there. This change touches only packages/client-runtime/src/state/ and the candidate registry.

Note: the agent ship-gate tooling (.githooks/, scripts/agent-pre-push.mjs, pnpm pr:ready) lives in fork/changes and does not exist on fork/candidates, so the equivalent checks above were run by hand.

[patroza]

Imports open upstream PR pingdotgg/t3code#5344 as one provenance commit on fork/candidates.

Source SHA: 783fd023c69f7ebc9c0ae657ce124cce77d94587 (upstream commits b623dc23b + 783fd023c, squashed).

Imported behavior

• reduceThreadStreamItems: a pure reducer that folds a batch of OrchestrationThreadStreamItems into one state and one persistable snapshot, preserving sequence dedup, completion markers, and deletion handling.
• Stream.groupedWithin(64, 16ms) on the live thread subscription, so a burst of thread events publishes the SubscriptionRefonce instead of once per event. Web and mobile stop rebuilding large thread views for every streamed event.
• eventBatchSize on EnvironmentThreadStateOptions, plus upstream's two regression tests: ordered single-publication bursts, and persisting a settled snapshot when a batch ends with a non-persistable turn start.

Local adaptations

• Kept our httpSnapshotLoadAttempted guard around the HTTP snapshot fallback (upstream does not have it); the call now goes through applyItems([...]).
• Restored setDeleted, which upstream deleted. Our terminal thread-deleted subscription failure (terminalSnapshotReason) never reaches the item stream, so it cannot go through the batch reducer. Cache removal is now shared between that path and the reducer path via removeCachedThread.

Excluded

• tasks/todo.md — the upstream author's scratch checklist.

Follow-up (belongs to fork/changes, not this layer)

fork/changes adds reloadFromServer and the result.kind === "reload-required" branch on top of setThread, which this import removes. Rebasing fork/changes onto this layer will conflict there (deliberately — it must not be resolved by dropping the branch). The intended re-expression: have the reducer report the reload-required index, commit state up to it, run reloadFromServer(), then re-enter applyItems with the remaining items — with reloadFromServer itself calling applyItems([{ kind: "snapshot", snapshot: fresh.value }]) instead of the deleted setThread.

Risks worth knowing

• Live state is now coalesced up to 16 ms (~1 frame). Under TestClock, a batch smaller than 64 items only flushes when the clock advances; upstream's harness pins eventBatchSize: 1 so existing tests keep their per-item assertions.
• applyItems snapshots state / awaitingCompletion at batch start and writes them back at the end, so a concurrent setDisconnected / setStreamError from the supervisor fiber can be clobbered. The race is pre-existing, but the window widens from per-event to ~16 ms.

Validation

• vp test run packages/client-runtime/src/state/{threads-sync,threadReducer,threads-atoms}.test.ts — 52 passed
• vp run -F @t3tools/client-runtime -F @t3tools/web test — 210 files / 1874 tests passed
• vp check — 0 errors
• vp run -F @t3tools/client-runtime -F @t3tools/contracts -F @t3tools/web -F @t3tools/mobile typecheck — clean

Workspace-wide typecheck is not meaningful locally on this layer: this worktree's node_modules was installed from fork/changes (@anthropic-ai/claude-agent-sdk@^0.3.220) while fork/candidates pins ^0.3.170, which produces unrelated ClaudeAdapter type errors in apps/server. Fork CI installs from this branch's own lockfile, so it is the authority there. This change touches only packages/client-runtime/src/state/ and the candidate registry.

Note: the agent ship-gate tooling (.githooks/, scripts/agent-pre-push.mjs, pnpm pr:ready) lives in fork/changes and does not exist on fork/candidates, so the equivalent checks above were run by hand.

[patroza]

Imports open upstream PR pingdotgg/t3code#5344 as one provenance commit on fork/candidates.

Source SHA: 783fd023c69f7ebc9c0ae657ce124cce77d94587 (upstream commits b623dc23b + 783fd023c, squashed).

Imported behavior

• reduceThreadStreamItems: a pure reducer that folds a batch of OrchestrationThreadStreamItems into one state and one persistable snapshot, preserving sequence dedup, completion markers, and deletion handling.
• Stream.groupedWithin(64, 16ms) on the live thread subscription, so a burst of thread events publishes the SubscriptionRefonce instead of once per event. Web and mobile stop rebuilding large thread views for every streamed event.
• eventBatchSize on EnvironmentThreadStateOptions, plus upstream's two regression tests: ordered single-publication bursts, and persisting a settled snapshot when a batch ends with a non-persistable turn start.

Local adaptations

• Kept our httpSnapshotLoadAttempted guard around the HTTP snapshot fallback (upstream does not have it); the call now goes through applyItems([...]).
• Restored setDeleted, which upstream deleted. Our terminal thread-deleted subscription failure (terminalSnapshotReason) never reaches the item stream, so it cannot go through the batch reducer. Cache removal is now shared between that path and the reducer path via removeCachedThread.

Excluded

• tasks/todo.md — the upstream author's scratch checklist.

Follow-up (belongs to fork/changes, not this layer)

fork/changes adds reloadFromServer and the result.kind === "reload-required" branch on top of setThread, which this import removes. Rebasing fork/changes onto this layer will conflict there (deliberately — it must not be resolved by dropping the branch). The intended re-expression: have the reducer report the reload-required index, commit state up to it, run reloadFromServer(), then re-enter applyItems with the remaining items — with reloadFromServer itself calling applyItems([{ kind: "snapshot", snapshot: fresh.value }]) instead of the deleted setThread.

Risks worth knowing

• Live state is now coalesced up to 16 ms (~1 frame). Under TestClock, a batch smaller than 64 items only flushes when the clock advances; upstream's harness pins eventBatchSize: 1 so existing tests keep their per-item assertions.
• applyItems snapshots state / awaitingCompletion at batch start and writes them back at the end, so a concurrent setDisconnected / setStreamError from the supervisor fiber can be clobbered. The race is pre-existing, but the window widens from per-event to ~16 ms.

Validation

• vp test run packages/client-runtime/src/state/{threads-sync,threadReducer,threads-atoms}.test.ts — 52 passed
• vp run -F @t3tools/client-runtime -F @t3tools/web test — 210 files / 1874 tests passed
• vp check — 0 errors
• vp run -F @t3tools/client-runtime -F @t3tools/contracts -F @t3tools/web -F @t3tools/mobile typecheck — clean

Workspace-wide typecheck is not meaningful locally on this layer: this worktree's node_modules was installed from fork/changes (@anthropic-ai/claude-agent-sdk@^0.3.220) while fork/candidates pins ^0.3.170, which produces unrelated ClaudeAdapter type errors in apps/server. Fork CI installs from this branch's own lockfile, so it is the authority there. This change touches only packages/client-runtime/src/state/ and the candidate registry.

Note: the agent ship-gate tooling (.githooks/, scripts/agent-pre-push.mjs, pnpm pr:ready) lives in fork/changes and does not exist on fork/candidates, so the equivalent checks above were run by hand.