gtdb_ground.py default paths miss the KG-Hub kg-microbe layout, and strain-level ids no longer ground #746

Two separate findings from grounding a new record on 2026-09-07. Both are real
and reproducible; neither is urgent.

1. The default kg-microbe search list misses a layout that exists here

just ground-taxa-gtdb --ncbi-id NCBITaxon:1488 fails with:

[gtdb] NCBI2GTDB mapping not found. Tried:
  .../ontology/Mechs/kg-microbe/data/raw/NCBI2GTDB.tsv.gz
  .../ontology/kg-microbe/data/raw/NCBI2GTDB.tsv.gz
  .../VIMSS/kg-microbe/data/raw/NCBI2GTDB.tsv.gz

The file is present, at
.../ontology/KG-Hub/KG-Microbe/kg-microbe/data/raw/NCBI2GTDB.tsv.gz — the same
parent directory that holds the culturebotai-claw checkout the CLAW_SRC
default also misses (../../culturebotai-claw/src does not resolve here
either, so just new-history needs CLAW_SRC= passed explicitly).

Passing --kg-microbe-dir works, so this is only a default. Two options:
add the KG-Hub/KG-Microbe/ sibling to the search list (consistent with the
existing three guesses), or drop the guessing and document KG_MICROBE_DIR.
The same question applies to CLAW_SRC.

Found with an ignore-independent find over ~/Documents/VIMSS, so the file's
absence from the three default paths is established, not inferred from a
gitignore-respecting search.

2. Strain-level NCBITaxon ids do not ground against the current mapping

With the mapping found, species rank grounds cleanly:

Strain rank does not, by id or by name:

Both ids resolve fine in NCBITaxon (checked against the local OAK build:
"Clostridium acetobutylicum ATCC 824", "Clostridium ljungdahlii DSM 13528").

This matters because the KB already stores a GROUNDED block for one of them.
kb/communities/Clostridium_Ecoli_Nitratidesulfovibrio_Minority_Mediator_Consortium.yaml
carries ncbi_source_id: NCBITaxon:272562, total_genomes: 2, and
mapping_source: kg-microbe NCBI2GTDB.tsv.gz; GTDB release latest (built 2026-07-25)
— a block the tool no longer reproduces from the mapping now on disk
(sha256:761d5537a7edb4b0, same 2026-07-25 mtime).

So one of these is true and I did not establish which:

• the mapping changed between the run that wrote that block and now, despite the
  matching build date; or
• the strain-rank code path changed (exclude_unnamed since Compare the two GTDB denominators, and make the named-species filter the default (#371, #375) #372/A third denominator option: exclude sp./uncultured MAG rows, which fixes Acetobacter under both rules #375 is the
  obvious suspect — a strain row is not a binomial); or
• that block was written by an older tool vintage and is now unreproducible.

The third would mean stored groundings can silently drift out of reproducibility
with no gate noticing, which is the same shape as #387 but for provenance rather
than coherence. Worth a check: re-derive every stored gtdb_classification from
the current mapping and count how many still reproduce.

Meanwhile the workaround is to ground at species rank and carry the strain in
strain_designation, which is what CommunityMech:000336 does.