feat: surface estimated token cost + require confirmation before launching resource-intensive skills

Problem

When a skill like deep-research fans out 15–20 subagents, it can consume a significant portion of a session's token quota with no prior warning. In my case, triggering it mid-session while actively coding in another context consumed ~25% of my quota before I could intervene and kill it.

Proposed behavior

• Skills that spawn more than ~5 agents (or exceed a configurable token threshold) should surface an estimated cost before executing — e.g. "This will spawn ~18 agents and use approximately 200k–400k tokens (~25% of a standard session quota). Proceed?"
• User must explicitly confirm before the workflow launches
• Ideally configurable per-skill or globally via a setting like resource_warnings: true

Why this matters

Users who run Claude Code across multiple concurrent sessions or contexts have no way to anticipate quota impact before a workflow fires. A one-line warning before launch would prevent accidental exhaustion entirely.

Acceptance criteria

• [ ] Resource-intensive skills display estimated agent count and token range before executing
• [ ] User sees a confirmation prompt and can decline without any agents being spawned
• [ ] Behavior is configurable (opt-out for power users who always want to proceed)