docs: add upstream-sync playbook for agents

[tqtensor]

Summary

• Adds a root-level CLAUDE.md so any Claude Code session opened in this repo gets the upstream-sync workflow on entry.
• Codifies the rules that emerged while resolving Upstream sync: 5 new commit(s) from perplexityai/modelcontextprotocol #1: don't merge/push/close without per-step approval, prefer rebase-merge, ship via PR (never push to main), no AI attribution.
• Captures established content filters: skip upstream's install badges and npm badge, keep the separate VS Code section with servers / type: stdio, adapt commands to github:tqtensor/... + OPENROUTER_API_KEY.
• Documents the gh pr create from a fork pitfall (defaults to parent repo unless --repo and --head are passed) so future PRs don't accidentally land in perplexityai/modelcontextprotocol.

Test plan

• CLAUDE.md renders correctly on GitHub.
• Open this repo in a fresh Claude Code session and confirm the playbook is loaded into context (Claude should be able to summarize the upstream-sync workflow without re-reading the file).
• Next time .github/workflows/upstream-sync.yml opens a new sync issue, an agent following this playbook produces a single-commit PR against the fork (not upstream), with Closes #<n> in the body.