fix: report failed transfers instead of ending in success

Upload and download failures were printed and then forgotten: a batch
that lost files still exited 0, and a file the client could not write
was listed as if it had arrived. Nothing downstream could tell.

- Return a verdict from every transfer worker and raise once at the
  end, so a run that lost a file exits 2.
- Raise the error when a downloaded file cannot be written locally,
  instead of printing it and reporting the path as a success.
- Write a download beside its destination and move it into place once
  complete, and refuse a destination that cannot be written, so a
  failed transfer no longer leaves a truncated file behind.
- Carry on through the remaining sub-folders when a recursive download
  loses a file or cannot create a folder locally.
- Name the file and give the reason in every failure message.
- Count a failure the API layer did not raise, such as a file that
  disappeared between the directory walk and its turn to be sent.
- Extract the directory walk from `Uploader.upload` to keep it within
  the complexity limit.
- Cover all of the above in `tests/test_transfer.py`.

CHANGELOG.md

@@ -5,6 +5,13 @@ All notable changes to this project will be documented in this file.
 ## [Unreleased]
 
 ### Fixed
+- Reported a failed upload to the caller. A file the server refused was printed and then forgotten, so a batch that lost files still ended in success and scripts could not tell.
+- Reported a failed single-file download to the caller, which was counted internally but never raised, so only recursive downloads ever ended in failure.
+- Raised the error when a downloaded file could not be written locally. A permission error was printed and the path was then listed as if the file had arrived.
+- Wrote a download to a temporary file beside its destination and moved it into place once the whole body had arrived. A transfer that failed part way used to leave a truncated file under the real name, and a destination that could not be opened was then deleted even though nothing had been written to it. A destination the client may not write is now refused before anything is fetched.
+- Named the file in the message when an upload failed, and counted a failure the API layer did not raise, such as a file that disappeared between the directory walk and its turn to be sent.
+- Carried on through the remaining sub-folders when a recursive download lost a file or could not be created locally, instead of abandoning the rest of the tree at the first failure. Every failure is now printed against its own file and reported once at the end.
+- Said why a download failed, rather than printing the path alone.
 - Serialised the token refresh across processes. Concurrent `mdrs` invocations shared one refresh token and each sent it, so a provider that rotates refresh tokens accepted the first and refused the rest.
 - Wrote the login cache through a temporary file so a reader can no longer catch it mid-truncation and discard the session.
 - Joined the base URL and the API's relative path correctly in `ls --json` output, which produced a doubled separator when the configured URL ended with one.