Add idea: decouple staging from execution for agent code-sandboxes

[rkoster]

What is this note about?

A spark prompted by digging into Heroku's code-execution sandbox implementation: its
dependency resolution happens live per call (pip install into a throwaway venv against
public PyPI), which doesn't fit enterprise CF deployments running offline/mirrored
buildpacks. This idea proposes splitting "stage an environment from a dependency manifest"
from "execute in an already-staged sandbox" — modeled on CF's own staging/droplet split —
as two decoupled primitives (e.g. MCP tools), so the platform can cache/reuse staged
environments and avoid re-resolving dependencies on every sandbox invocation.

Checklist

• My idea lives in ideas/ and uses a lowercase kebab-case filename.
• It has a title (YAML title:).