Majority-of-genomes grounding can contradict GTDB's type-species rule

[realmarcin]

Found reviewing PR #372. The general form of #373.

gtdb_ground.py grounds a higher-rank NCBI taxon to whichever GTDB taxon holds the most genomes. GTDB, however, assigns names by nomenclatural type: the lineage containing the type species keeps the unsuffixed name, and the rest take alphabetic suffixes.

These rules disagree whenever a non-type clade is more heavily sequenced. Verified against GTDB R226 for all five taxa where the two denominators differ:

taxon	grounded to	type species	type sits in
Enterococcus	g__Enterococcus_B	E. faecalis	g__Enterococcus
Pseudomonas	g__Pseudomonas	P. aeruginosa	g__Pseudomonas ✓
Acetobacter	g__Acetobacter	A. aceti	g__Acetobacter ✓
Leptospirillum	g__Leptospirillum	L. ferrooxidans	g__Leptospirillum ✓
Bacillus	g__Bacillus	B. subtilis	g__Bacillus ✓

Four of five agree by luck rather than by construction — the majority happens to be the type clade. Only Enterococcus diverges today, but nothing prevents the others from flipping as sequencing depth shifts.

Options:

1. Prefer the type-anchored clade when the NCBI name has an unsuffixed GTDB counterpart — arguably what "grounding NCBI genus X" should mean.
2. Flag the disagreement rather than resolve it: emit the majority answer but mark that it is not the type clade, so a curator sees it.
3. Leave as-is and treat Enterococcus is grounded to a non-type GTDB clade: g__Enterococcus_B is the faecium clade #373 as a one-off.

Option 2 is cheap and loses nothing; option 1 changes semantics and deserves a decision.

[realmarcin]

Concrete evidence for this, from investigating #377. Two live KB instances where the genome-weighted vote contradicts the type-species rule:

taxon	vote	type-bearing split	records
Enterococcus (NCBITaxon:1350)	g__Enterococcus_B 0.598	g__Enterococcus 0.348 — holds E. faecalis	1
Bacillota (NCBITaxon:1239)	p__Bacillota_A 0.501	p__Bacillota 0.410 — holds Bacillus	4

Verified against the crosswalk: g__Bacillus sits in p__Bacillota, g__Clostridium in p__Bacillota_A; E. faecalis sits in g__Enterococcus. So in both cases the winner is the most-sequenced clade and the loser is the one the name belongs to.

The blocker is representational, which is worth knowing before anyone picks this up. Storing the type-bearing term means storing a minority majority_fraction, and gtdb_coherence rejects < 0.5 as fraction_out_of_range — rightly, since the field means "the fraction backing the winner". Attempting it fails test_every_stored_block_is_coherent, test_the_known_near_ties_are_still_there and two others.

So this issue cannot be closed by re-grounding. It needs one of: AMBIGUOUS + gtdb_candidates (schema-conformant now, and used by the sibling Clostridium entry in the very same record for its 45 splits); relaxing the [0.5, 1.0] bound for curated: true blocks; or a separate slot for the nomenclatural basis leaving majority_fraction alone.

Note the Bacillota four are also the near-tie population tests/test_gtdb_near_tie_marker.py pins for #396, so the three issues have to move together.