add a rapids-singlecell analysis skill

[Intron7]

Summary

• bundle a model-neutral rapids-singlecell analysis skill with the Python distribution
• add installers for Codex, Claude, Claude Science, AGENTS-style clients, and explicit destinations
• add disposable GPU/kernel preflight plus live public-API search and description helpers
• make public facade exports explicit so the generated API map can distinguish supported and deprecated symbols
• document installation, memory, Dask, notebook, spatial, and ecosystem-interoperability workflows
• add semantic regression tests for scientific and operational guidance

Why

Cold agentic runs could complete large RSC analyses, but repeatedly lost expensive iterations to GPU layer residency, API discovery, notebook execution, plotting scale, cluster/niche validation, and package-boundary decisions. Shipping the skill with RSC keeps those instructions version-matched to the installed package and gives agents a live API contract instead of a static function catalog.

User impact

A pip-installed release now carries the matching skill and exposes:

• rapids-singlecell-install-skills
• rapids-singlecell-check-kernel
• python -m rapids_singlecell_skills.api

The skill defaults supported computation to RSC, preserves explicit ecosystem boundaries, requires human-style executed notebooks and reports, and keeps inferential decisions tied to the biological replication unit.

Validation

• pytest tests/test_install_skills.py -q — 28 passed
• Ruff checks and repository pre-commit hooks passed after formatter updates
• skill structure validator passed
• live API hand layer validated against 59 public symbols
• live searches for squidpy, pertpy, and decoupler resolved to public RSC facade symbols
• three independent cold-agent workflow checks passed
• uv build --wheel completed successfully
• built wheel contains the full skill bundle and both console entry points

[Zethson]

Cool! I really think the dask guide has a lot of value because I had the impression LLMs don't know how to use dask well.

The usual - having one line makes markdown readable. My eyes are going crazy.

IDK whether you wrote a single word here. If it's mostly agent generated, an agent would have probably known all of this already?

Could we extract a lot of the "how to do analysis" stuff into a scverse wide skill? It's a waste of we all write a form of this in all of the scverse packages CC @PauBadiaM Same for the whole api stuff. Can we have a general version of that in scverse-misc? And installation, and kernel, ...
It's important to get this PR merged quickly but I'd that you @Intron7 and @PauBadiaM find ways to DRY this across scverse.