Why does a noisy sensor that repeats its previous bit after warming up worry a randomness assessor more than a noisy sensor whose output just looks messier?

Answer: Because copied bits create hidden dependencies that make the next bit easier to predict.

Because messy-looking output shows more variation, so the messier plot is the larger risk. — A messy plot can reassure the eye, but visual clutter is not an entropy assessment. NIST separates output testing from entropy-source assessment because a source can have memory or bias that lowers worst-case unpredictability. The messier-looking sensor may be less risky than the one with hidden dependencies.

Because warming adds physical entropy, so the repeated bit is probably harmless. — Warmth can be part of a physical noise source, but a warmer sensor is not automatically more random. If every bit partly copies the previous one, the extra heat has not removed the dependency; it has left a memory pattern that can lower worst-case unpredictability. The idea that heat adds entropy is only half right and misses what the assessor is testing.

Because copied bits create hidden dependencies that make the next bit easier to predict. ✓ — Repeating the previous bit is a dependency: the next value is partly constrained by the last one, so the worst-case guess can improve. NIST SP 800-90B treats real entropy sources as often non-IID, and predictor or Markov-style tests look for exactly this kind of memory. A stream can look noisy and still leak patterns an attacker could use.