What is actually happening. The corpus is a fixed 1080-cell grid of letters, laid out once and never changed. The alphabet is 27 symbols: space plus a–z. The embedding is a 27×27 bigram table counted from that exact grid — row i is P(next letter | current letter). Nothing is learned at runtime and nothing is random unless you switch pick to sample.
One step, in order. Read the letter under the agent → map it to a token id → fetch its row from the table → score each of the 8 neighbouring cells by the probability of that neighbour's letter → move to the highest. Candidates light red before the move; the path stays blue behind it.
Why it gets stuck. The table has 212 nonzero cells of 729. Most letter pairs never occur, so most neighbours score zero and the agent settles into short loops around high-frequency letters — e, t, a, space. That is not a bug in the walker. It is what a sparse transition table looks like when something has to navigate it.