A vanished child is absent

The one sentence

When a directory's child disappears while the directory is being listed, the child is absent. That is not a reason to fail the whole listing. FileSystemStorageAdapter.ListChildPaths takes a snapshot of a directory's subdirectories and then probes each one for content. A subdirectory that was deleted between the snapshot and its probe made the probe throw DirectoryNotFoundException, and that exception failed the listing of the PARENT.

What it cost

A failed listing is not loud. StorageAdapterMeshQueryProvider's walk catches it, records a dropped read and answers with nothing under that parent. What happens next depends on which read failed:

The Initial read. The frame says it is incomplete (SnapshotIncomplete). The synced-query cache then refuses to keep it, so the next GetQuery call asks again (The query fan-in's Initial gate).
A live re-query. Nothing marks it. ProcessBatch compares the empty answer with the rows the query already holds and emits Removed for every one of them. The write that triggered the re-query is never Added. Nothing triggers another re-query, so the live result stays wrong until some later write lands under the same parent.

The measured case is in MeshWeaver.Plugins, where AgentListOneChainTest.OneChain_RacingACreate_NeverRegressesFromHasToLacks failed in CI job 121872335269. A test class runs a mesh whose file-system persistence is the test binary's TestData folder, and the test classes that run beside it, four at a time, use the same folder. One iteration created an agent under ACME/Agent while a parallel class deleted a sibling directory there. The live query's re-query listed ACME/Agent, the probe of the deleted sibling threw, and the query never listed the new agent within its 30 s budget. A fresh chain listed it at once.

A local stress run reproduces it: the same race loop, with a thread that keeps creating and deleting sibling directories under ACME/Agent, missed 9 of 60 creates. The emissions show both shapes: n=174 has=False | n=0 has=False (a live re-query wiped the list), and n=0 as the very first frame (an incomplete Initial). With no churn the same loop missed 0 of 150, and with the fix it misses none with churn.

The deleting side does not need another process. A node delete prunes its own directory (Moving Nodes), so one mesh can delete a directory under a parent that one of its own live queries is listing at that moment.

The fix

ClassifyChildDirectory probes one child. A child that is gone by the time it is probed is Vanished and is left out of the listing, which is the same answer the listing would have given a moment later. A parent that disappears after its existence check lists as empty, which is the same as a parent that was never there. Only DirectoryNotFoundException is treated this way. Every other fault (permissions, I/O) still fails the listing and is still recorded as a dropped read.

Regression test

FileSystemListingToleratesAVanishedChildTest (MeshWeaver.Hosting.Test) creates a child directory, deletes it, and then probes it, which is the step that decides the interleaving. Without the fix the probe throws DirectoryNotFoundException, and with it the probe answers Vanished. The negative control checks that a present child is still classified by its content (Node, Container, NoContent), so the test cannot pass by answering Vanished for everything. The adapter's own listing is not driven, because the snapshot and the probe run back to back inside one call and there is no seam to open the window between them.

What this does not claim