feat(ingest, backend): improve ingest curation warning - add diff

[anna-parker]

resolves #3084

Adds the option to filter by AccessionVersion on the get-unprocessed-metadata endpoint. Have ingest call this endpoint to get all previous metadata (includes sequence hashes, so it will indirectly tell us if a sequence has been changed) when we get into the case where ingest would like to revise a curated sequence.

Ingest will then add a diff of the curated metadata and the metadata it would like to revise to the notification so that curators can decide better to to handle the situation.

PR Checklist

• All necessary documentation has been adapted for backend endpoints
• The backend endpoint updates are covered by unit tests
• Any manual testing that has been done is documented (i.e. what exactly was tested?)

1. Ingest sequences from the INSDC with initial hash
2. Curate a sequence from the INSDC as a superuser (add a revision, do not change hash)
3. Ingest ignores revision as the hash has not changed
4. Log into database and manually modify the hash of the curated entry (unprocessed data not original data)
5. Run the ingest cronjob and verify that an informative message is sent to the loculus notifications slack:

🚀 Preview: https://curation-warning-add-diff.loculus.org

[claude]

Claude finished @anna-parker's task in 1m 56s —— View job

---

Code Review

• Gather context and read changed files
• Review backend changes (Kotlin)
• Review ingest changes (Python)
• Post inline comments
• Final summary

---

Good feature overall — adding the metadata diff to the curation warning makes the Slack notification much more actionable. The backend filter implementation is clean and well-tested. A couple of issues worth addressing before merge:

One functional bug: get_sequence_status() is called before the early-return in get_submitted when output is None, so every call from calculate_metadata_diff fetches statuses for all sequences in the group unnecessarily. See inline comment on loculus_client.py:498-501. Fix this →

Code style: Two print() calls in calculate_metadata_diff (lines 127, 131) should be logger.debug() or logger.info(). See inline comment.

Slack message length: The full metadata diff is inlined directly into the notification string. For sequences with many metadata fields, this could easily exceed Slack's ~3000 char limit, causing the notification to be silently dropped. Worth truncating or summarising. See inline comment on compare_hashes.py:181.

Other observations:

• The backend accession version parsing using lastIndexOf('.') is the right approach — it correctly handles accessions that themselves contain dots.
• The two new backend tests cover the happy path and the invalid-format 400 case well.
• No automated tests for calculate_metadata_diff — acceptable given it's a notification path and was manually verified, but worth a note.
• The # type: ignore is acknowledged in an inline comment; the structural match is correct at runtime.