gtdb_ground.py results are batch-sensitive: the same taxon grounds differently per-record vs whole-KB

[realmarcin]

Found reviewing PR #364. Pre-existing.

collect_rows() breaks after the first matching rank column, so which rank a mapping row is indexed under depends on which other taxa are in the same run. The same taxon can therefore resolve differently depending on batch composition.

Demonstrated with Euryarchaeota (NCBITaxon:28890, in High_Solids_Switchgrass_Methanogenic_Microbiome.yaml):

• ungrounded-only batch → AMBIGUOUS (no block emitted)
• whole-KB batch → grounds to GTDB:p__Halobacteriota

This matters twice over:

1. The remediation advice in Ground the taxon left stale by an id edit, and pin grounding to its source id (#314) #364's new test is batch-sensitive. Its failure message says "re-run gtdb_ground.py --apply on these records" — a per-record run may not reproduce what a whole-KB run produced, or vice versa.
2. Grounding is not reproducible from the record alone, which undercuts mapping_source as provenance: the same mapping release and the same id can yield different answers.

Fix: make collect_rows() index a row under every rank it matches rather than breaking at the first, so resolution depends only on the taxon and the mapping table.