test(sftp): third-party client interop validation for the raised channel sizing and pipelined write path (FileZilla, WinSCP, sshj/Cyberduck)

[inureyes]

Problem / Background

PR #224 (closing #187) raised maximum_packet_size to 65535 and window_size to 8 MiB, and restructured the server-side SFTP write path (bounded in-flight requests, sequential write coalescing, deferred flush). OpenSSH sftp and scp (both protocol modes) were verified locally with byte-for-byte integrity, but the interop matrix from #187's acceptance criteria (FileZilla, WinSCP, an sshj-based client such as Cyberduck) was not exercised in the implementation environment. That remaining acceptance criterion moves to this issue.

Motivation

These clients have historically surfaced bssh-server regressions: see #215, where sshj/Cyberduck and sftp -C sessions dropped mid-transfer. Channel window/packet-size negotiation and response ordering under pipelining are exactly the surfaces that third-party client stacks stress differently from OpenSSH, so the raised sizing and the pipelined write path need explicit interop coverage.

Tasks / Acceptance Criteria

• 1 GiB upload and download with FileZilla against bssh-server, byte-for-byte integrity verified. (Waived: GUI-only client and no display on the validation host; a PuTTY-core psftp proxy run is the suggested follow-up. See the closing comment.)
• 1 GiB upload and download with WinSCP against bssh-server, byte-for-byte integrity verified. (Waived: Windows-only GUI client; same PuTTY-core proxy note applies. See the closing comment.)
• 1 GiB upload and download with an sshj-based client (e.g. Cyberduck) against bssh-server, byte-for-byte integrity verified.
• Confirm negotiated packet/window sizes are sane per client (debug logs) and no mid-transfer stalls or drops occur.
• Results table documented (client, direction, throughput, integrity, notes).

Technical Considerations

Server compression remains off by default (#220 added an opt-in server.compression option); interop runs should use the default configuration.

Related

• Follow-up to perf(sftp): bssh-server throughput ~50% of OpenSSH from 32 KiB channel fragmentation, serial server-side writes, and per-packet copies #187, implemented by PR perf(sftp): raise channel sizing, pipeline server writes, tune release build #224.
• Interop precedent: fix(sftp): sshj/Cyberduck clients dropped on initial REALPATH/STAT via russh strict-parser SshEncoding::Length on incoming decode #215.

[inureyes]

Interop validation results (closing summary)

Environment: bssh-server built from main at e4f2810 (post-#224), loopback 127.0.0.1, all files on tmpfs, 2 GiB random file (exceeds the 1 GiB criterion), byte-for-byte cmp after every transfer, server pinned to a single core, default server config (compression off).

Results table

Client	Upload	Download	Integrity	Negotiation / notes
OpenSSH 9.6p1 sftp	460 MiB/s avg	655 MiB/s avg	OK	measured during the #225 evaluation on the same setup; no stalls
sshj 0.38.0 (the stack under Cyberduck)	224 MiB/s	220 MiB/s	OK	negotiated kex curve25519-sha256, hostkey ssh-ed25519, cipher chacha20-poly1305@openssh.com both directions, compression none/none; server advertised max packet 65535 and an 8 MiB window, and the client consumed data in 65535-byte packets (transport and connection debug logs); ls/stat/rm exercised; no stalls or drops across 2 GiB each way
paramiko 2.12.0 (supplementary, not in the original criteria)	OK (8 MiB in 0.2 s)	hangs with default prefetch	fail	pre-existing server-side bug, not a #224 regression (reproduces identically on fbbf226); the same script completes in 0.1 s against OpenSSH sshd on the same host; prefetch=False completes but at ~37 ms per request; tracked in #227

sshj throughput is client-CPU-bound (Java client, unpinned); the relevant results for this issue are correctness, negotiation sanity, and stall-freedom, all of which pass.

FileZilla and WinSCP

Waived on this pass: both are GUI-only (no headless mode) and the validation host has no display; WinSCP is additionally Windows-only. Both use PuTTY-derived SSH transports, so a psftp (putty-tools) run is the natural headless proxy when an environment allows it. The interop harness used here is being added to the repository under tools/bench/ to make that follow-up cheap.

Disposition

Closing as completed per maintainer decision. The sshj criterion, the negotiation checks, and the results table are satisfied; FileZilla/WinSCP are waived as above; the paramiko finding is tracked separately in #227.