v0.3.1

Bug Fixes

• Fix --think silently accepting invalid values — Typos like --think medum now error instead of silently disabling thinking. Valid values: true, low, medium, high.
• Fix --num-predict 0 sending Ollama default instead of unlimited — 0 now correctly sends -1 to Ollama for unlimited generation.
• Reject negative --num-predict and --seed — Previously silently fell back to Ollama defaults; now returns a clear validation error.
• Reject --dry-run in CLI mode — Previously accepted but silently ignored; now errors with guidance to use TUI mode.
• Normalize --think to lowercase — --think=HIGH now works correctly instead of falling through to disabled.
• Distinguish "not installed" vs "not running" in connection errors — Shows platform-specific install instructions when Ollama binary isn't found, scoped to localhost URLs only.
• Surface Ollama error messages on non-200 responses — Error messages from Ollama (e.g., "model does not support thinking") are now shown instead of generic HTTP status errors.
• Guard BuildSchedule against empty model slice — Prevents a panic in balanced multi-round mode with no models.

Improvements

• IPv4-only test server helper for better CI portability
• 16 new regression tests covering flag validation, request serialization, and error paths

Full Changelog: v0.3.0...v0.3.1