Keep sync filesystem work off the runtime

Author: myobie

CHANGELOG.md

@@ -8,6 +8,13 @@ EXPERIMENTAL, so on-disk formats and the CLI may change without notice.
 
 ### Fixed
 
+- **A sync folder walk no longer stalls unrelated Fabric streams.** Scan,
+  materialization, JSON encoding, metadata checks, and persistence now run on
+  Tokio's bounded blocking pool without holding the live node or disk locks.
+  In a one-worker, five-second pipe window with repeated 500 ms walks, maximum
+  delivery delay fell from 1.195 seconds to 16.9 ms. The source maximum stayed
+  below 10.5 ms in both windows.
+
 - **Fabric commands no longer replace the formatting in `peers.toml`.** Add,
   remove, policy, and Git remote changes update only the affected entries.
   User comments, spacing, order, unknown fields, and an edited or removed file

README.md

@@ -965,6 +965,9 @@ opening or reading files does not schedule sync work. Write bursts settle for
 watcher-driven sync per two-second window. Inbound no-op sessions already queued
 for the same durable folder generation reuse its pre-merge scan, and routine
 sync-accept path snapshots are sampled in the default validation log.
+Folder scans, materialization, and state persistence run on Tokio's bounded
+blocking pool. Fabric holds the live node and disk locks only while it snapshots
+or applies a completed phase, so filesystem latency does not stop other streams.
 
 Entries live in an authoritative, hand-editable `syncs.toml` next to `peers.toml`
 (`~/.config/fabric/syncs.toml` for the default home, `<home>/syncs.toml` with