Scan four resolutions in the guided tutorial, and band the result

[shanikawm]

Closes the last item from the domain-expert review — comment 4, "I typically
run 3 resolutions at the same time and compare them before deciding which one to
use for downstream analysis."

#93 gave find_clusters the vector form. This uses it.

The sweep

Both sides now run [0.4, 0.8, 1.2, 0.5] — find_clusters(pbmc, resolution=[...]) and FindClusters(pbmc, resolution = c(...)) — and every
resolution is scored against R:

resolution	truecell	Seurat	ARI	concordance
0.4	9	9	0.8958	0.9602
0.5	8	9	0.8987	0.9549
0.8	11	11	0.8264	0.9174
1.2	12	12	0.7995	0.8647

Two things a single resolution could not show. The cluster count matches
exactly at 0.4, 0.8 and 1.2 — so the 8-vs-9 split this tutorial has always led
with is specific to resolution 0.5, not a standing property of the port. And
agreement falls as resolution rises, which is what finer partitions should do:
more boundaries, more chances for two Louvain runs to land in different optima.

0.5 is given last on purpose. Seurat leaves the object on the last
resolution in the sequence, and UMAP, the markers, the annotation and the
handoff CSV are all written against it. Verified rather than assumed: the
tutorial's four outputs are byte-identical to what main produces. A test
pins the ordering, since reordering the list re-points all of them silently.

Found while doing it: a stale headline number

The tutorial documented ARI 0.938 while measuring 0.899 — across six files
(tutorials/README.md ×2, pbmc3k_tutorial.md ×2, docs/fidelity.md,
docs/quickstart.md), with the concordance figures stale to match: 2,554/2,638
cells and 0.968, against a measured 2,519 and 0.955.

I checked whether this PR caused it. It did not — R's vector-form FindClusters
produces a partition identical to the single 0.5 call, confirmed by running
the main script and diffing per cell.

The likely cause is the graph fixes in #67–#71 moving cells between clusters —
the same drift the DE tutorial's deseq2 top50 band did catch at the time
(25 → 22). This tutorial had no band on its headline number, so it went stale in
six documents instead of failing once. Every swept resolution now carries a
declared band, checked by --report.

The [0.9.0] changelog entry quoting 0.938 is left alone — it records what was
measured then.

Verification

Five mutants, all caught: reordering the sweep so it no longer ends on 0.5;
desynchronising the two languages' resolution lists; dropping a band; widening
one to vacuity; and reverting R to a single-resolution call. The band check was
separately confirmed to fire by narrowing one until it failed.

1179 passed, 25 skipped. mypy at its 4-error baseline. --report exits zero.
A changed ridge-plot PNG was reverted as rendering churn, the underlying
clusters having been verified identical.

Also

Generalises the CSV-precision lint from #94: dtype=str parses no floats, so it
cannot misparse one. That was a per-file exemption for the DE hex reader and is
now a rule — which is what caught this PR's own new reads.

🤖 Generated with Claude Code