Add VERCEL_PLUGIN_DISABLED_SKILLS env var to opt out of specific skills

Problem

The plugin ships ~30 skills, some of which (like next-forge) match aggressively on common patterns — @repo/, keys.ts, env.ts, proxy.ts, monorepo, pnpm-workspace, etc. In repos that aren't next-forge but happen to use similar conventions, the skill fires repeatedly through both UserPromptSubmit lexical matching and SessionStart import/path detection.

Currently the only way to suppress a single skill is to delete its directory from the plugin cache (and patch generated/skill-manifest.json), which gets clobbered on every plugin update.

Proposed

An env var like VERCEL_PLUGIN_DISABLED_SKILLS=next-forge,chat-sdk that:

• excludes those skills from user-prompt-submit-skill-inject.mjs candidate matching
• excludes them from session-start-profiler.mjs suggestion mapping
• excludes their context chunks in vercel-context.mjs

It would mirror the existing VERCEL_PLUGIN_LIKELY_SKILLS shape (comma-separated skill names), just inverted.

Use case

Yarn 4 monorepo unrelated to next-forge that was triggering the skill on proxy.ts edits, monorepo discussion, and keys.ts references — none of which are actionable for that project.

Plugin version

vercel@0.40.0