GLM-5.2 documented S396's 20x context-sensitivity: the same writer in the same session produced a 12:1 compression ratio in testimony documents but 0.59:1 in methodology documents — providing strong evidence that compression grammar shifts with document purpose rather than being a fixed trait