Add agent-aware workspace launch

[svvayyy]

What changed

• add pagent launch adapters for Codex, Claude Code, and OpenClaw
• inject the committed Parallel Agent workflow as the agent's initial context
• add pagent instructions for ChatGPT, remote agents, and unsupported clients
• replace the raw cd && agent hint after workspace creation with managed launch guidance
• document the workflow and OpenClaw configuration caveat
• add integration and adapter tests

Why

Agents launched manually from a Parallel Agent worktree correctly received Git isolation, but still followed the repository's ordinary package-manager and development commands. That bypassed Pagent's port allocation, Vercel Development environment injection, process tracking, logs, and cleanup.

The new launch handoff makes the operational contract explicit without creating or modifying AGENTS.md, CLAUDE.md, or other repository instruction files.

Validation

• pnpm format:check
• pnpm typecheck
• pnpm test — 14 tests passed
• npm pack --dry-run
• real pagent launch ux-test --dry-run preview against the NetworkAudit test clone
• changed-file secret and instruction-file scan