install_skills.py: the failed-hash rollback runs skills remove without -a, so it can delete another agent's same-named skill #405

Finding (from the GPT-6 review of #391, 2026-09-27)

tools/adoption/install_skills.py installs each pinned skill for two agents only, but its rollback removes the skill for every agent:

• line 176: add_args = ["add", skill["url"], "--skill", name, "-g", "-y", "-a", "claude-code", "codex"]
• line 196, the rollback when installed content does not match the pinned hash or tree: run_skills_bin(skills_bin, ["remove", name, "-g", "-y"], ...)

In skills v1.7.0, remove without -a/--agent selects every known agent and recursively removes the matching skill directories (upstream src/remove.ts#L209, confirmed against the pinned 1.7.0 cli.mjs removeCommand/parseRemoveOptions). A failed-hash rollback can therefore delete a same-named skill that another agent owns (for example Cursor's), which this installer never installed. That breaks adoption/lifecycle.md:100: "Never uninstall all user tools to roll back one package."

Suggested fix: scope the rollback exactly like the install: ["remove", name, "-g", "-y", "-a", "claude-code", "codex"]. -a must come last, because the parser consumes the trailing words as agent names. Add a unit test that asserts remove carries the same -a list as add.

Related: #391 (its rollback text now uses the scoped form), docs/decisions/2026-09-25-skills-trial-and-usage.md:43.