Remove smbj: replace SMB file copy with a transfer through the WinRM command shell

[bertysentry]

Goal

Remove com.hierynomus:smbj — the last runtime dependency of winrm-java — and replace the SMB-based temporary-share file copy with a file transfer performed through the WinRM command shell itself, the way dependency-free clients like pywinrm (Copy-Item helpers), winrs-based tooling, and Ansible's win_copy do it.

This is the final step of the zero-dependency goal of the winrm-light effort (#103–#108) and the actual fix for #116:

com.hierynomus:smbj:0.14.0          0.58 MB
├─ org.bouncycastle:bcprov-jdk18on  8.18 MB   (90% of the ~9 MB standalone CLI JAR, #112)
├─ org.slf4j:slf4j-api              0.06 MB   (#116)
├─ net.engio:mbassador              0.09 MB
└─ com.hierynomus:asn-one           0.06 MB

smbj cannot run on stock JDK crypto (it needs MD4 for NTLM, AES-CMAC for SMB 3.x signing, AES-CCM for SMB 3.x encryption — none of which are in SunJCE), which is why it drags in BouncyCastle. Removing smbj removes all of the above in one move and takes the standalone JAR from ~9 MB to a few hundred KB.

Proposed design

Transfer the file over the already-authenticated, already-encrypted WinRM channel — no new port (445), no separate SMB authentication, works through firewalls that only allow 5985/5986:

1. Base64-encode the local file and split it into chunks that fit comfortably within the WinRM MaxEnvelopeSize (pywinrm/Ansible use ~1 KB–8 KB per command leg; we can size chunks dynamically from the negotiated envelope size).
2. Append each chunk on the remote side via the command shell, e.g. echo <chunk> >> C:\Windows\Temp\<name>.b64.tmp, or preferably a single PowerShell invocation that reads chunks from stdin to reduce round-trips.
3. Decode once at the end: certutil -decode or PowerShell [IO.File]::WriteAllBytes(..., [Convert]::FromBase64String((Get-Content ...))), then delete the temp .b64 file.
4. Verify integrity by comparing a hash (certutil -hashfile / Get-FileHash SHA-256) against the locally computed digest; fail with a WindowsRemoteException on mismatch.
5. Clean up the remote temp artifacts on failure paths too.

Performance note: base64-over-SOAP is slower than SMB for large payloads. That is acceptable — the existing SmbTempShare/WindowsTempShare use case is copying small script files for execution, not bulk data. Document the expected use (small files) and consider a size guard with a clear error message.

Scope

• Reimplement the SmbTempShare functionality (temporary remote directory + file copy + cleanup) on top of WindowsRemoteExecutor.executeCommand(...); keep the public API (WindowsTempShare contract) as stable as possible.
• Delete the smbj dependency from pom.xml.
• The standalone CLI JAR (Add a WinRM CLI and standalone executable JAR #112) shrinks accordingly with no shade-plugin exclusions needed.

Acceptance criteria

• mvn dependency:tree shows zero runtime/compile dependencies (test scope only).
• No org.bouncycastle, org.slf4j, net.engio, or com.hierynomus classes in the standalone JAR; JAR size back to a few hundred KB.
• File copy works live against a real Windows host over NTLM/HTTP (encrypted) and HTTPS, including: file with multibyte UTF-8 content, binary file, empty file, and a file large enough to require multiple chunks.
• Hash verification failure and mid-transfer errors surface as exceptions (no logging, per Remove SLF4J: the library must not depend on any logging framework #116) and leave no orphaned temp files on the remote host.
• Unit tests cover chunking, encoding, and failure paths; live coverage added to the existing live-test rig.

Closes the dependency-removal half of #116 (SLF4J) as a side effect.