settingSources: [] still loads user-level MCP servers and skills

Bug Description

When using the Claude Agent SDK with settingSources: [] (documented as "SDK isolation mode"), user-level MCP servers and skills are still loaded into agent sessions. This causes significant token overhead (~30K extra tokens per session) from tool definitions that the agent doesn't need.

Reproduction

import { query } from '@anthropic-ai/claude-agent-sdk';

for await (const msg of query({
  prompt: 'Say ok',
  options: {
    systemPrompt: 'Respond with just "ok".',
    model: 'claude-haiku-4-5',
    tools: [],
    settingSources: [],  // should mean full isolation
    maxTurns: 1,
    permissionMode: 'default',
  },
})) {
  // Check usage on assistant message — cache_creation_input_tokens is ~35K
  // even though no tools or settings should be loaded
}

Expected Behavior

With settingSources: [], the SDK docs state: "When omitted or empty, no filesystem settings are loaded (SDK isolation mode)." This should mean:

• No user-level MCP servers loaded (from ~/.claude/settings.json)
• No user-level skills loaded (from ~/.claude/skills/)
• No managed skills loaded (from /Library/Application Support/ClaudeCode/.claude/skills)

Actual Behavior

SDK debug logs show:

Loading skills from: managed=/Library/Application Support/ClaudeCode/.claude/skills, user=/Users/xxx/.claude/skills, project=[...]

And user-level MCP servers (Gmail, PostHog, Granola, Playwright, Linear — whatever the user has configured) are loaded, adding their tool definitions to the context.

Token impact: ~35K cache_creation tokens with settingSources: [] + tools: [], vs an expected baseline of ~5-10K for just the system prompt.

Workarounds Found

1. strictMcpConfig: true — blocks user MCP servers (reduces ~57K → ~26K tokens)
2. extraArgs: { 'disable-slash-commands': null } — should block user/managed skills (untested reduction)

Environment

• SDK version: 0.2.63
• Node: 24.13.0
• OS: macOS (Darwin 24.6.0, arm64)

Related Issues

• #176 — settingSources ignored in V2 Session API (similar symptom)
• #124 — Feature request for Tool Search to reduce token usage