[daltschu22]

Dsync could return success while transfers were still running or had failed, omit files during chunking, and write/delete testfile.dsync even during cloud dry runs. This change waits for bounded parallel transfers, reports failures, preserves filenames and dry-run behavior, and checks cloud source lists on the host performing the copy.

The implementation remains a single script organized around Fpart, Rsync, Rclone, and FilesystemOps, with a direct path-check → chunk → transfer workflow. Backend commands stay with their tools; scheduling, safe argument handling, and cancellation are shared. Popen supplies the process handles needed for concurrency and cleanup.

• Use argument lists and quoted SSH commands; send chunk lists over stdin instead of requiring shared working files. Preserve remote path spelling. Both SSH hops use batch authentication, no terminal, preserved stdin, a 15-second connection timeout, and 15-second keepalives with three missed replies allowed.
• Honor --rclone-config, remove cloud write probes, and fix recursive cloud basic chunking. Rclone's missing-file-list behavior can silently skip source entries; a small Python preflight on each transfer host validates the list and passes it to rclone through an unlinked temporary file. Disable interactive config-password prompts.
• Include hidden files and special names. Use NUL-delimited rsync lists and raw rclone lists; retain rsync symlinks and empty directories.
• Replace incompatible fpart -Z usage, detect traversal diagnostics even when fpart exits zero, and avoid recursive re-copying of fpart chunks.
• Stream basic chunking, cap open chunk files at 32, avoid eagerly collecting filenames from wide cloud source directories, skip empty jobs, and enforce -n during reuse.
• Handle SIGINT/SIGTERM and process-registration races; finish cleanup despite additional signals and stop surviving local process-group members after the grace period.
• Lock working directories, stage partitions, validate reuse, and replace only verified unchanged chunks from an ownership manifest. Preserve unrelated/modified files and reject overlapping local source/destination/work paths.
• Add documentation and a Python 3.10/3.14 CI matrix.

Validation:

• 44 regression/integration tests passed locally with no skips, using Python 3.14.4, rsync 3.4.1, fpart 1.7.0, and rclone 1.60.1-DEV. Coverage includes actual local copies, special filenames, dry runs, permissions, reuse, ownership, concurrency, locking, interruption, missing cloud sources on workers, a dropped loopback HTTP connection, and 128 chunks under a 64-descriptor limit. Python compilation and git diff --check passed.
• Manual validation against an isolated real loopback OpenSSH server passed: rsync on source workers, rsync to destination hosts, nested worker-to-destination SSH, and rclone on workers. This exercises the real SSH transport; it does not establish cross-machine NFS behavior.
• Real rclone uploads to a temporary loopback WebDAV server passed. Stopping the server during a throttled 16 MiB upload returned failure; restarting it and rerunning with --reuse produced the complete verified file.
• A local 100,000-file, eight-chunk cloud-list generation comparison reduced peak Python allocation from 6.735 MiB to 1.011 MiB (tracemalloc); elapsed time was 2.554 s versus 2.412 s. This is a chunk-generation measurement, not a cloud/NFS throughput benchmark.

Compatibility and limits: legacy chunks need regeneration. Modified/unrecognized working files require review or a new working directory. Cloud workers now require python3, writable temporary space for file lists, and noninteractive rclone credentials. Source checks add metadata lookups and cannot prevent source changes after validation; use a stable snapshot. New files require regenerated lists. Newline/carriage-return filenames fail explicitly for older rclone compatibility. Host files support DNS/IPv4/SSH aliases; SSH filesystem destinations must be absolute. Actual multi-machine NFS mounts and external cloud providers remain untested. SSH keepalives do not impose a total job deadline or resolve blocked NFS I/O; remote cleanup depends on SSH/backend disconnect behavior. Failures can leave completed chunks, and incomplete-upload cleanup/resumption is backend-specific. The README explains resource scaling and conditional rclone destination-listing tuning.