Document execution safety boundaries

[koriyoshi2041]

Summary

• documents the read / construct / broadcast split in the skill catalog
• clarifies that broadcast is the execution boundary where real funds can move
• clarifies that spend_limit only caps paid skill-call fees, not transaction notional, slippage, liquidity, bridge, or token exposure
• suggests an agent-integration policy: construct first, show exact unsigned tx or steps, then broadcast only after explicit approval

Verification

• npm run build
• git diff --check