It looks better because the anomaly score and likelihood were consistently erratic when the signal was calm. HTM systems contain entropy. You can’t get rid of it. But if you dial it in like this, you can find anomalies like you just did!
Is that graphic not what you expected? This looks like a good anomaly indication to me. In both cases:
You got a significant jump in the log likelihood when the amplitude changed, and again after a sudden spike and another amplitude change. You are in a place now where you should run some experiments, or even find some real world signal to parse.