fix: harden reasoning effort state handling

[ian-pascoe]

Summary

• Split reasoning-effort state into current, persisted, and temporary reset variants so idle resets happen only once.
• Validate reasoning effort levels before prompting or mutating state, and consolidate system guidance into a single prompt entry.
• Add regression tests covering reset behavior, persisted settings, invalid levels, and new-session guidance.

Verification

• pnpm run format:check
• pnpm run lint
• pnpm run test
• pnpm run typecheck
• pnpm run build