Allow safe mid-conversation model and effort changes

[ai-yappa]

Summary

• add provider-neutral reasoning effort levels to sandbox start and phase messages
• preserve the existing Pi conversation when changing model or effort
• reject model/effort changes during an active stream to avoid racing provider state
• apply effort changes to plan, refine, execute, and follow-up agent turns

Validation

• pnpm --filter @codevil/shared test ✅
• pnpm --filter @codevil/shared build ✅
• pnpm --filter @codevil/sandbox-image build ✅
• node --test packages/sandbox-image/test/pi-driver.test.mjs ✅
• Full sandbox test run was attempted; two pre-existing environment-sensitive runtime assertions fail because repository setup emits install/preview-app events in this workspace.