fix: default BUSCO lineage when predict cannot resolve taxonomy

[vinvomero]

Summary

Fixes #93. funannotate2 predict aborts with AttributeError: 'bool' object has no attribute 'get' when no taxonomy is available for the species: lookup_taxonomy() returns False on any JGI API failure (offline, 5xx, unrecognized name), and the training-data fallback params["taxonomy"] is also False for a novel/unpublished organism. The crash lands at the BUSCO-lineage selection step — after the ab initio predictions have already run — so a long run is lost to a late failure on exactly the "taxonomy unavailable" case.

While fixing this I also found a latent, separate issue on the same path: for a non-empty taxonomy that matches nothing in busco_taxonomy, best_taxonomy() falls through to random.choice(), so a random (but valid) BUSCO lineage is selected non-deterministically — and because the pick is always a valid key, a result in busco_taxonomy check can't catch it.

What this does

• Adds busco_lineage_from_taxonomy() in utilities.py — a small pure helper that resolves a valid lineage deterministically. It returns a default (fungi) unless the taxonomy genuinely overlaps the reference on superkingdom/kingdom, and otherwise returns the matched lineage. This closes both the crash and the random-lineage path.
• Routes predict() through the helper, defaulting to the fungi lineage and logging a clear warning that names the species and flags that a generic lineage is unreliable for a non-fungal genome (there is no --busco-lineage/--taxonomy override on predict, so it self-heals rather than aborting).
• Hardens the offline branch to params.get("taxonomy"), so a params file that predates the taxonomy key no longer KeyErrors before the guard runs.

Tests

Adds regression, determinism, and edge-case coverage in tests/unit/test_utilities_taxonomy.py: the #93 crash inputs (False/None), a per-kingdom determinism invariant (so a future busco_taxonomy edit can't silently reintroduce the random tie-break), and non-string / partial / exact-match cases. All taxonomy unit tests pass locally.

Note — possible follow-up

I scoped this PR to predict to match the issue. The same "taxonomy can be False" pattern also reaches choose_best_busco_species() in annotate.py and train.py (and choose_best_augustus_species() in train.py), so funannotate2 annotate / train can hit the same crash class on their default (no --busco-lineage / --augustus-species) path. Happy to send a focused follow-up routing those call sites through the same helper if you'd like it — kept out of this PR to keep it reviewable.

🤖 Generated with Claude Code