Fix missing UploadedFile causing silent release failures

[Goober5000]

The multiupload/finish idempotency path returned success without verifying the UploadedFile DB record still exists. If the record was lost (e.g. expired) after the upload completed, the release endpoint would reject the mod with "archive missing" — but clients like KNet never saw the error because create_release() returned HTTP 200 for all error responses.

Two fixes:

• multiupload/finish now checks for the UploadedFile record when ChunkedUpload.done is True, and re-creates it from the file on disk if missing. If the file is also gone, it resets done=False so the client can retry.
• create_release() error responses now return proper HTTP status codes (400/403/500) instead of 200.

Depends on #49; in draft until that is merged

[notimaginative]

The OK(200) issue I reported is a bug in KNet. It reports the status code on error rather than the error message it was sent.

The HTTP status code thing in fsnebula is definitely an issue, however it's not one that can be fixed, as KNet and old-knossos depend on the current behavior to work properly. Fixing the status codes on fsnebula would almost certainly break proper error handling on both clients.

[chief1983]

I followed up with Claude today about the thing I had a question about last night on Discord, about whether there was a risk of it flagging a file still in the temp location as successfully uploaded. Doesn't sound like there is an issue, but it did flag some other things:

However, there is a different subtle problem with the recovery code: it re-creates the UploadedFile with expires=-1 and the correct permanent path, but it doesn't verify the file's checksum. If the file on disk is corrupted or partial (e.g., disk issue), it would happily create a DB record pointing to a bad file. The original path does checksum verification at lines 182-185 before calling make_permanent().

Also, the recovery creates the record with content_checksum and vp_checksum from the retry request's form data, which should be the same as the original — but there's an implicit trust there without verification.

Summary: The temp-stuck-file scenario can't happen due to the ordering of make_permanent() before done=True. But the recovery path does skip checksum verification, which the normal path performs. That's the real gap in PR 50's upload.py changes.