Agent CRDT follower: a frame consumed while graph scope is unavailable is never projected until an unrelated later update

[christian-byrne]

Problem / Goal

Split out of the BLOCKER on #16353 (#16353 (comment)), which correctly identified this and is a test-only PR.

LayoutFollowerBridge.onDocUpdate merges the Yjs bytes and advances lastSeq before the ECS adapter is given the frame. EcsFollowerAdapter.applyFrame → GraphMutations.batch() returns false when getScope() is null, so nothing is projected. The bridge has no way to un-consume the bytes: the replica already mutated and the state vector already advanced (same behaviour proven in #16372).

AgentPanelRoot.vue's getScope() reads boundTabFor(workflowId)?.activeState?.id, while the subscription gate isBoundWorkflowActive only compares boundTabFor(bound)?.path === active.path. The two can disagree — a bound tab whose path matches the active workflow but whose changeTracker has not hydrated yet (or whose serialized workflow carries no id) is subscribed and scope-less at the same time.

Why there is no recovery. Reproduced locally against the real bridge, adapter and createGraphMutations:

subscribe #1 state vector: AA==            (empty)
frame seq=1 delivered, scope null          -> applyFrame() === false, graph empty
follower state vector now:  Acy648sPDA==   (already contains seq=1)
scope becomes available
bridge.resubscribe()
subscribe #2 state vector: Acy648sPDA==    (identical to the post-merge vector)
graph nodes after resubscribe: []

The resubscribe carries a vector that already covers the withheld update, so the host computes an empty delta and sends no catch-up. Nothing local re-drives projection either. EcsFollowerAdapter does keep session.reconcileNextFrame = true on a rejected batch, so the debt is retained — but only a further semantic frame discharges it:

frame seq=2 delivered, scope available
graph nodes: ["1", "2"]                    (the withheld node finally lands)

So the graph stays empty for as long as the session is idle. On a session that has finished its turn, that is indefinite.

Proposed Solution

A trigger is what is missing. The bytes are already local, so recovery needs no host round trip — only a re-run of the adapter's authoritative reconcile once scope exists. Two candidate shapes:

1. Explicit reprojection. Add EcsFollowerAdapter.retryProjection(workflowId) that replays the reconcile with the rejected frame's context, and call it when scope becomes available.
2. Gate the subscription on projectability. Fold activeState?.id != null into isBoundWorkflowActive so the follower never subscribes into a state it cannot project. The bridge would then hold sentWorkflowId === null and drop inbound frames before merging, so the later subscribe's state vector still misses them and the host's catch-up is correct.

Trap for whoever picks this up: the obvious trigger is not reactive. ComfyWorkflow.changeTracker is markRaw'd ("Non-reactive raw object" — src/platform/workflow/management/stores/comfyWorkflow.ts:43), so a watch/watchEffect on activeState?.id will never fire and would ship as a silently dead recovery path.

Acceptance Criteria

• A frame consumed while getScope() is null is projected once scope becomes available, with no further host frame and no resubscribe required.
• A test covers that path and fails when the trigger is removed (the trigger must be proven live, not merely present — see the markRaw trap above).
• No partial projection while scope is unavailable: the existing assertions in followerSeam.integration.test.ts that nothing is written and no layout op is emitted continue to hold.
• The existing retained-debt baseline in followerSeam.integration.test.ts (the withheld node is reconciled in by the next frame) still passes.