Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See