feat-upload-coordinator.md

Enhancement: Extract the upload state machine into a driver-agnostic coordinator

The upload state machine (TUS session management, postprocessing event loop, antivirus integration, and restart safety) has been extracted from decomposedfs into a new coordinator in pkg/upload. Every storage driver now inherits TUS chunked uploads, postprocessing, and AV scanning without reimplementing any of it.

Drivers integrate by implementing four new methods on the storage interface:

• MarkProcessing sets or clears a "processing" flag on a resource so readers see a grayed-out placeholder while bytes are in flight. Drivers that do not need concurrent-upload protection may implement this as a no-op.
• PrepareUpload is called after all bytes are received and before postprocessing begins. Decomposedfs uses this to lock the node, snapshot the previous version, and propagate the optimistic size change. Drivers with no such requirements may return immediately.
• CommitUpload writes the staged bytes to the resource and receives pre-computed checksums.
• RollbackUpload is the inverse of PrepareUpload and is called when postprocessing fails or is aborted. Drivers that returned immediately from PrepareUpload may return nil. The RollbackInfo struct carries the node identity from the upload session rather than from live node metadata, so a rollback can still release the quota of a node whose metadata has become unreadable (e.g. because an ancestor was trashed mid-upload).

The coordinator owns the upload session files for the decomposedfs driver at the same on-disk location as before (<root>/uploads/), so existing in-flight uploads continue without interruption and no migration is required.

Configuration:

Both storageprovider and dataprovider gain an upload_directory config key that sets the local directory where temporary upload session files and staged bytes are stored. For decomposedfs this is optional; the coordinator falls back to <root>/uploads/ inside the driver's own root directory. For drivers that have no local filesystem root, upload_directory must be set explicitly; otherwise the service fails to start.

The postprocessing consumer settings (asyncfileuploads, consumer_group, numconsumers, mount_id) are read from the driver's own config block, the same keys decomposedfs already uses. No new top-level config is introduced.

#702 #703 #714 #715 #717 #720 #721