[R1] A marker is RAISED here when `avg_log2FC > 0` AND `auc > 0.5`. Both, not either:
`avg_log2FC` is a MEAN over every cell in the group, including the ones that do not
express the gene, so a handful of extremely high cells can pull it positive even while
most of the distribution points the other way; `auc` is the rank-based version of the same
comparison and cannot be moved by a few cells the way a mean can. Where they disagree --
a positive `avg_log2FC` with `auc <= 0.5` -- trust `auc`: the gene is not raised, whatever
the mean says. `significant` is a stricter and separate statement built on the same two
numbers: the row clears all four of `avg_log2FC` above 0.25, `pct_in` above 10, `auc`
above 0.5, and an adjusted p-value below 0.05. A row can be raised without being
significant, and a row can be significant while carrying little weight for an identity
many cell types share the gene with.

[R1b] `auc` reads the whole expression distribution, not a detection threshold, so it
tells apart a gene truly concentrated at high levels in a MINORITY of the cluster's cells
(auc still above 0.5, because most in-cells outrank most out-cells even though few of them
individually cross a detection cutoff) from a gene whose apparent rise is really a handful
of outliers dragging a mean upward while the bulk of the cluster ranks no higher than the
rest of the tissue (auc at or below 0.5). The second shape is not evidence for the
identity, however large its `avg_log2FC`.

[R2] An unraised marker means one of two different things, and `pct_in` tells them
apart. Unraised with `pct_in` near zero is evidence AGAINST the identity: the cells do
not express what the resource says defines it. Unraised with a high `pct_in` and a
similarly high `pct_out` is not evidence against it; a cell type that makes up much of
the dataset drives its own defining markers to a flat fold change, because the
comparison is against the rest of the same tissue.

[R3] Panel rows are ordered by `auc`, the most separating row first. `n_pub` is a fact
on each row, not its rank: it says how often a pairing has been REPORTED, it does not rank
identities against each other, and it is not a measure of how essential a gene is to an
identity -- a heavily reported marker is frequently one the neighbouring cell types also
express, which is why `specificity` is a separate column and why the most-published
marker of an identity is often its most famous rather than its most reliable. Weigh each
gene by what it does in these cells and by what its sentence actually claims for it.

[R3b] `n_pub` = 1 is a single-study pairing: one paper once wrote the gene next to the
cell-type name. Read such a row, and any sentence marked `single_study`, as one study's
observation rather than as the field's definition, and give it correspondingly little
weight against a replicated row that points the other way. It is not nothing -- a
replicated panel is corroborated, not contradicted, by a single-study marker that agrees
with it -- but an identity, a refinement or a qualifier resting on single-study pairings
alone rests on one laboratory's finding.

[R3c] What a sentence says matters as much as that it exists. A gene the literature uses
to DEFINE an identity separates it. A gene appearing only as a control, as one comparison
inside a single analysis, as a marker of a different state of the same lineage, or in a
list of things the authors also measured, does not separate it -- whatever its detection
here, and whatever its publication count.
