Move the docs stacklet to Paperless-ngx 3.0.4

[arthware-dev]

Closes #50.

Moves docs from Paperless-ngx 2.20.15 to 3.0.4. A watchtower pull of :latest once made this jump by accident and broke document filing; this is the deliberate version, with the two things that actually break fixed.

Duplicate detection was broken twice over

3.0 changed duplicate handling in two independent ways, and unmodified we would have hit both.

It stops rejecting duplicates at all. Rejection is now conditional on PAPERLESS_CONSUMER_DELETE_DUPLICATES, which defaults off. With it unset, re-sending the same letter files a second copy and reports success. Now set in the compose.

The rejection no longer carries a message. _DUPLICATE_RE scraped the twin's name out of the failure text. On 3.x there is no text, so wait_task returned None and the archivist reported a generic upload failure instead of "already filed". It now reads result_data.duplicate_of and looks up the title, with the 2.x text path kept for older images.

Worth knowing for reviewers: on a rejected upload 3.x sets related_document_ids to the twin. Nothing but result_data separates that payload from a successful filing.

A second search word narrows again

3.0 also replaced Whoosh with tantivy, whose parser joins bare terms with OR. Since we wildcard every token too, homer car insurance matched 18 of 24 documents in the demo archive: one word was enough to hit. Joining with AND fixes it, measured on the same archive:

user types	before	after
marge recipe	11	1
lisa report card	6	1

Queries where the caller wrote their own operator or a quoted phrase keep their structure, because injecting AND into either breaks it: "birth certificate" would become "birth AND certificate" and match nothing, and insurance OR recipe would become a parse error.

This trades one failure mode for another. homer car insurance now returns 0, because that document is titled "Auto Insurance Policy 2026" and the word "car" is nowhere in it. Ranking and a confidence cutoff are tracked separately in FAM-18, along with the fact that the German compound rationale for the wildcard is now unverified (the demo set is English only).

On method

The first version of the duplicate parser was written from the upstream source at v3.0.4, and the real server disagreed with it. Source reading said a duplicate task ends success with no related documents; a live 3.0.4 returns failure with the twin's id present. The offline tests passed against the wrong fixture. What is pinned now was captured by uploading the same file twice and dumping /api/tasks/ verbatim, and TestTaskApiShape in the new e2e keeps it honest by handing the real payload to the real parser.

Two of my own e2e assertions counted documents through search, which reads an asynchronously updated index. They passed alone and failed inside the full suite. They now ask the document list instead.

Also here

• Upgrade note. Paperless migrates its database on first 3.x start and 2.x will not boot afterwards. Anyone already rolled forward by watchtower cannot pin back without a backup taken before that boot. The admin guide now says so, and says what to back up first.
• stacktests down <stacklet> never worked. exec cannot run a shell function, so an explicit target died with exec: stack_cli: not found and tore nothing down. The no-argument form masked it.
• AGENTS.md non-negotiable 11. Seed instances through the front door. Loading documents straight into Paperless skips the archivist, so the corpus has no tags, correspondent, type, summary or vault entry, and any measurement taken on it describes a system we do not ship. Learned the hard way in this branch.