Follow-ups from the Azimuth integration: vocabulary seeding, upstream workaround, CL-id cross-validation

[Marius1311]

Low priority — two small follow-ups from #43, plus one upstream watch.

1. Seed a level's vocabulary from Azimuth (cta markers assemble --ref-label-csv)

cta markers assemble can already seed markers_<level>.json from any labelled reference via
from_reference() (it runs rank_genes_groups over an obs column). Azimuth's per-cell labels
make the query itself such a reference — azimuth_medium's 85 pan-body classes are a
plausible L2 prior for a broad or multi-tissue human query, and they arrive with a CL crosswalk.

Today that takes three manual steps. Making it one needs roughly 15 lines:

• add --percell-out PATH to cta crosscheck azimuth (~5 lines) so the per-cell labels are
  written, not just the per-cluster aggregate
• add --ref-label-csv PATH to assemble_markers.py: read the CSV, join on obs_names, then
  fall through to the existing from_reference path

The output is a prior to curate, exactly like the CellTypist source — assemble_markers.py
already says so in its docstring, and that applies here unchanged. Worth doing only if L2
hierarchy drafting turns out to be a bottleneck in practice.

2. Remove the "False" workaround if upstream fixes it

src/cta/run_azimuth.py works around
satijalab/panhumanpy#47: panhumanpy writes
the literal string "False" into azimuth_broad/medium/fine for cells whose 8 hierarchical
heads do not agree on a self-consistent path. We substitute final_level_labels for those cells
and report the rate as frac_inconsistent_hierarchy — which is useful on its own as an
out-of-distribution alarm, so only the substitution would go away, not the metric.

The issue is open with no maintainer response. If they introduce a proper sentinel (None/NaN)
or a documented fallback, revisit _resolve in run_azimuth.py.

3. CL-id cross-validation at write-back — decided AGAINST for now

Recorded so it isn't re-proposed without new evidence. cta crosscheck azimuth writes
azimuth_cl_<level>.csv, an independent label → CL:ID assertion, and it is tempting to have
cta report write-back warn when it disagrees with the cl_id Claude assigned.

Not worth building yet: Claude's labels and Azimuth's come from different vocabularies, so a
mismatch is the expected case, not the exception — a warning would fire on most rows and erode
a gate that currently has a clean signal-to-noise ratio. Only a minority of the shipped crosswalk
rows are skos:exactMatch; restricting to those is what would make it viable at all.