TaxonMech is the tenth fleet member and carries a card, a graph node, a stat line and a manifest badge on /mechs/. It has no heatmap row and no overlap chords, because it is absent from RECORD_GLOBS in scripts/fleet/roots.py, so prefix_census.py and build_subsets.py never measure it.
The page states the limitation in two places, so nothing published is false. But TaxonMech is the largest corpus in the fleet and its absence from the shared-vocabulary layer makes the heatmap misleading about what the fleet holds in common — NCBITaxon in particular, which is exactly what TaxonMech is built on.
What it takes
scripts/fleet/mech_stats.py already carries the glob, in EXTRA_GLOBS:
EXTRA_GLOBS = {"TaxonMech": ["data/taxa/**/*.yaml"]}Moving it into roots.RECORD_GLOBS is the whole change. The cost is the scan: 625,960 records, half again the size of ProteinTraitsMech at 429,291, across two passes that each read every file. Budget 15–20 minutes for a full pipeline run afterwards, and check that build_subsets.py's per-cell record caps still produce sensible index sizes at that volume — assets/fleet/ is already ~5 MB.
Once measured, drop the two "covers nine of the ten Mechs" caveats: _fleet/fleet_fragment.html (the heat note, which substitutes <!--FLEET_CENSUS_COUNT-->, and the graph legend, where "nine" is typed by hand) and _fleet/README.md.
TaxonMech is the tenth fleet member and carries a card, a graph node, a stat line and a manifest badge on
/mechs/. It has no heatmap row and no overlap chords, because it is absent fromRECORD_GLOBSinscripts/fleet/roots.py, soprefix_census.pyandbuild_subsets.pynever measure it.The page states the limitation in two places, so nothing published is false. But TaxonMech is the largest corpus in the fleet and its absence from the shared-vocabulary layer makes the heatmap misleading about what the fleet holds in common — NCBITaxon in particular, which is exactly what TaxonMech is built on.
What it takes
scripts/fleet/mech_stats.pyalready carries the glob, inEXTRA_GLOBS:Moving it into
roots.RECORD_GLOBSis the whole change. The cost is the scan: 625,960 records, half again the size of ProteinTraitsMech at 429,291, across two passes that each read every file. Budget 15–20 minutes for a full pipeline run afterwards, and check thatbuild_subsets.py's per-cell record caps still produce sensible index sizes at that volume —assets/fleet/is already ~5 MB.Once measured, drop the two "covers nine of the ten Mechs" caveats:
_fleet/fleet_fragment.html(the heat note, which substitutes<!--FLEET_CENSUS_COUNT-->, and the graph legend, where "nine" is typed by hand) and_fleet/README.md.