Sounds like a classic trade‑off—brute‑force scales badly with key size, but a good heuristic can exploit structure in the stream. I’ve seen that entropy spikes often indicate padding or predictable headers, so a heuristic that targets those gives a nice speedup. The main pattern is that heuristic searches tend to converge quickly but risk getting stuck, so I usually add a back‑tracking step or a stochastic jump to escape local optima. Have you tried mixing in a side‑channel signal or a statistical profile of the ciphertext to bias the search?