docs(skill): build-once → run frozen for scheduler/HPC envs

[Marius1311]

Closes #35.

Problem

On scheduler/HPC clusters the skill's prescribed pixi run cta … workflow makes pixi solve + validate against one shared env prefix at runtime, on every call. Observed failure modes: the dependency solve OOM/SIGSEGVs on memory-capped head/login nodes; concurrent runs serialize on the per-call exclusive prefix lock (even cta … --help queues behind a heavy lock-holder); slow first invocations from re-resolution.

Fix (skill-side, scheduler-general)

The skill ships a committed pixi.lock, so runtime never needs to solve. This PR makes that the default and documents the scheduler path:

• --frozen everywhere. pixi install --frozen / pixi run --frozen cta … is now the single prescribed invocation on every profile — runtime uses the locked env as-is and never solves. A one-line dev escape-hatch notes when to drop --frozen (intentional dependency changes, in a separate clone).
• Build once, bypass the per-call lock (scheduler). SKILL.md §1c and a new "Build once, run frozen — the scheduler / HPC reality" subsection in reference/compute_environments.md cover: don't solve on a capped head node; build once where there's memory; for concurrent/batch use resolve the prefix once (CTA=$(pixi run --frozen which cta)) and call "$CTA" … directly; don't mutate shared env state mid-run; clear stale local .pixi/envs when using detached-environments.
• Consistency. Updated the remaining prescribed-command spots (SKILL.md commands table, tool_registry.md bootstrap row, scheduler-handoff example).
• No cluster specifics — phrased purely in generic scheduler terms (memory-capped head node + one shared, lockfile-frozen prefix); explicitly no module systems, proxies, or concrete paths.

Verification (off a head node)

Ran on a login node (eu-login-42):

• pixi install --frozen → 74 MB peak RSS, exit 0 — links the committed env, does not run a memory-hungry solve.
• pixi run --frozen cta --help → exit 0.
• CTA=$(pixi run --frozen which cta) resolves and "$CTA" --help runs — the direct-bin concurrency path works.

Scope

Skill repo only. The cluster-specific operator steps (prebuild before fan-out, concrete detached-env scratch path, per-session prompt addendum) are tracked separately for quadbio/celltype-anno-bench docs/running-the-skill.md, per the issue.

🤖 Generated with Claude Code