Duplicate skills load from repository and installed plugin

Problem

Copilot CLI loads the same skill twice when a repository contains a project skill and an enabled user plugin contributes the identical skill. This doubles the skill description surface in repositories that develop their own plugin.

Observed with Copilot CLI 1.0.79-9 in rjmurillo/ai-agents, where .claude/skills is the canonical source for the installed project-toolkit plugin.

Reproduction

cd <repo-that-contains-the-plugin-source>
copilot plugins list --json | jq -r '\n[.plugins[] | select(.kind=="skill") | {name,scope,source,description:(.description // "")}]\n| group_by(.name)[]\n| select((map(.scope)|index("repository")) and (map(.scope)|index("plugin")))\n| [. [0].name, (length|tostring), ([.[].scope+":"+.[].source]|join(","))]\n| @tsv'\n```\n\nMeasured result:\n\n```text\nduplicate_names=97\nidentical_descriptions=97\nduplicate_description_bytes=40624\n```\n\nExample name:\n\n```text\ncontext-optimizer  plugin  plugin\ncontext-optimizer  repository  project\n```\n\nThe installed plugin is enabled at user scope. No repository-scoped setting was found to disable that one plugin only in its source repository.\n\n## Expected\n\nWhen a repository skill and plugin skill have the same canonical name and identical content, load one entry. Project scope could take precedence, or the CLI could deduplicate by name plus content hash. A repository-scoped plugin-disable override would also provide a workaround.\n\n## Control\n\nThis is specific to skills. A canary placed in `plugin/instructions/*.instructions.md` did not load through `--plugin-dir`, while a repository `.github/instructions` canary did load.\n

Beginne damit, das Problem mit copilot plugins list --json in einem Repository nachzustellen, das .claude/skills und ein aktiviertes Plugin enthält, und verfolge anschließend, wie Repository- und Plugin-Skills erkannt werden. Vergleiche die Einträge repository und plugin hinsichtlich des kanonischen Namens und des Inhalts und überprüfe, dass identische Skills einen einzigen geladenen Eintrag erzeugen, ohne die dokumentierten Instruction-Canaries zu beeinflussen.