fix(node): keep Schrödinger clusters across re-interview

[Apollon77]

Problem

A client cluster present in local storage but absent from the peer's descriptor serverList ("Schrödinger's cluster") survived startup (loadCache) but was deleted on the first live wildcard read after a fresh data migration. A server restart re-loaded it correctly — the deletion only struck the live read path.

Observed in matterjs-server (see home-assistant/addons#4668): a migrated node's EP1 carries data for manufacturer cluster 0x125dfc11 while its descriptor serverList omits it; the first re-interview logged Removing ...ep1.neoCluster from active endpoint.

Root cause

Asymmetry between loadCache and live mutate:

• loadCache processes the descriptor before injecting the endpoint's cluster behaviors, so the Schrödinger cluster is never in currentlySupported and is kept.
• In mutate, the endpoint is already active. When the descriptor chunk (server list without the cluster) arrives before the cluster's attribute data, #synchronizeDescriptor schedules the cluster for deletion. The cancel-pending-delete logic lived only inside #synchronizeCluster's if (cluster.behavior === undefined) branch — unreachable for an already-active cluster — so the later attribute data never cleared the deletion and #rebuild dropped it.

Storage is not truly erased, so the next loadCache restores the cluster, which is why a restart "fixed" it.

Fix

ClientStructure:

• New #clustersWithDataThisInteraction set + #preserveAbsentCluster helper (clears pendingDelete, warns once).
• Populate the set and call the helper in #updateCluster and the applyWireChanges update branch → cancels a deletion a descriptor scheduled earlier in the same interaction (descriptor-first ordering).
• #synchronizeDescriptor skips scheduling pendingDelete for a cluster already in the set → handles data-first ordering.
• The set is cleared at the start of mutate/applyWireChanges (exception-safe; no cross-interaction leak).

Live mutate is now symmetric with loadCache: a cluster the peer still serves data for survives a re-interview regardless of chunk order.

Test

New ClientNodeTest case "Schrödinger's cluster": establishes the cluster active-but-absent-from-serverList in one interaction, then re-interviews with descriptor-first ordering. Reproduces the exact deletion before the fix; passes after.

Verification

• Full packages/node suite: 1251/1251 (ESM/CJS), 1242/1242 (Web)
• format-verify clean, lint clean

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes a client-structure edge case in @matter/node where a “Schrödinger” cluster (present in cached/local state and still reported by the peer via attributes, but omitted from the peer’s Descriptor serverList) could be incorrectly deleted during the first live re-interview due to chunk ordering.

Changes:

• Add per-interaction tracking of clusters that received data, and use it to prevent/suppress descriptor-driven deletion scheduling for those clusters regardless of report ordering.
• Centralize “preserve absent cluster” logic to cancel pending deletions when attribute/wire data arrives for a cluster that was previously marked for deletion.
• Add a regression test that reproduces descriptor-first ordering and asserts the cluster survives re-interview.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
packages/node/src/node/client/ClientStructure.ts	Track clusters receiving data during an interaction and prevent descriptor sync from deleting clusters that still receive attribute/wire updates.
packages/node/test/node/ClientNodeTest.ts	Add a regression test for descriptor-first re-interview ordering to ensure “Schrödinger” clusters persist.

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request