fix(ingest): regroup revoke needs to work when number of sequences is greater than batch size, add tests

[anna-parker]

If more than batch_chunk_size records to regroup the new sequences are already POSTed to Loculus by the time the KeyError fires, and it fires before the revoke loop, so zero old accessions get revoked → the mis-grouped originals stay live alongside the new ones. Worse failure than just erroring out early.

Screenshot

PR Checklist

• All necessary documentation has been adapted.
• The implemented feature is covered by appropriate, automated tests.
• Any manual testing that has been done is documented (i.e. what exactly was tested?)

🚀 Preview: Add preview label to enable

[claude]

Claude finished @anna-parker's task in 3m 13s —— View job

---

Claude Code Review

• Gather context (read diff + surrounding code)
• Review loculus_client.py changes
• Review new tests
• Post review

The core fix is correct and the diagnosis in the description is right: post_fasta_batches previously kept only the last batch's response, so submission_id_to_new_accessions was missing every submissionId outside the final chunk, and the KeyError fired after the new sequences were already POSTed. Accumulating across batches and pushing .json() down into post_fasta_batches (so submit_or_revise no longer has to unwrap a possibly-None response) is the right shape. Converting the bare json.load(open(...)) to a context manager is a good drive-by.

Four inline comments, in rough priority order:

1. Test helper writes one base per line (test_regroup_and_revoke.py:82) — extend on a string iterates characters. Harmless but unintended and slow. Fix this →
2. Abort message is misleading (loculus_client.py:395-404) — the submit has already happened, so aborting here is the "old and new both live" state, not a prevention of it. Either revoke the resolvable accessions before raising, or reword so the operator knows manual cleanup is needed.
3. Batching test doesn't assert batching (test_regroup_and_revoke.py:160-169) — it would still pass if batch_chunk_size were ignored, and the fake backend ignores sequenceFile, so metadata/fasta misalignment across a batch boundary isn't covered.
4. Unbounded result accumulation (loculus_client.py:266) — one dict per record is now retained for the whole run, including in the submit/revise path where call_loculus.py:92 discards it. Not a blocker, cheap to bound.

Nothing security-relevant; make_request already raises on non-OK, so failed batches propagate rather than silently producing a malformed results list.

The three tests are well constructed otherwise — building on the real CCHF pipeline output rather than hand-rolled fixtures means they exercise the actual add_seq_to_batch join, the import style matches test_merge_group_overrides.py, and the relative paths line up with how CI invokes pytest tests/ from ingest/.

Note: I was not able to run pytest or ruff check in this environment (commands outside the permitted set), so the review is static only. ingest/ isn't covered by the ruff-format pre-commit hook (it's scoped to ena-submission/ and preprocessing/nextclade/) and ingest-tests.yaml only runs snakefmt, so a few blocks in the new test file are formatted more loosely than ruff format would produce — harmless, just flagging in case you run it locally.
· branch fix_regroup_revoke