[BUG] claude.ai marketplace MCP server instructions inject into context on every session regardless of Tool Search

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Problem
Tool Search defers MCP tool schemas, but MCP server instruction blocks (system prompt injections from claude.ai marketplace servers) load unconditionally on every session start. These are separate mechanisms, and Tool Search does not affect the instruction context.

With several Claude. ai marketplace servers configured (e.g., Figma, Slack), instruction injection alone accounts for 30–55% of context consumed on a fresh window — before any user input.

What Should Happen?

Expected: MCP server instructions load on demand or are scoped to sessions where the server is actually used
Actual: All instruction blocks inject on startup regardless of whether the server is needed

Error Messages/Logs

Steps to Reproduce

Steps to reproduce

1. Configure 3+ Claude.AI marketplace MCP servers
2. Open a fresh Claude Code window
3. Note the context percentage shown in the status bar before sending any message

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

v2.1.109

Platform

Claude Code CLI

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Related: #47645 covers the tool schema side of the same symptom — this is the
instruction/prompt side, which requires a separate fix.