Disabled plugins still inject context via SessionStart and UserPromptSubmit hooks

Plugins set to "enabledPlugins": false in ~/.claude/settings.json continue to fire their SessionStart and UserPromptSubmit hooks, injecting large amounts of context into the conversation window.

Steps to reproduce

1. Install a plugin with SessionStart hooks (e.g., vercel@claude-plugins-official)
2. Disable it: set "vercel@claude-plugins-official": false in enabledPlugins in ~/.claude/settings.json
3. Confirm it shows ✘ disabled via claude plugin list
4. Start a new session or run /clear

Observed behavior

• The plugin's SessionStart hook fires and injects its full context (in the case of vercel, a ~15KB knowledge graph)
• UserPromptSubmit hooks also fire on every message, matching keywords and demanding skill invocations
• This fills ~15-20% of the context window before the user says anything

Expected behavior

A disabled plugin should be completely silent — no hook execution, no context injection.

Environment

• Claude Code (CLI)
• macOS Darwin 23.0.0
• Plugin: vercel@claude-plugins-official (version 87b8d0f6bb0a)

Workaround

Full uninstall via claude plugin uninstall <name> — but this forces reinstall when the plugin is needed again, instead of a simple enable/disable toggle.