The context-sharing gap you describe is real, but there's a second failure mode in cascade setups that's quieter: tool-call schema drift between models. If the cheap model calls a function and the cascade kicks in mid-task because the answer looked weak, the fallback model doesn't just lose conversation context, it can also structure tool calls differently (argument names, nesting, even whether it emits a call at all for the same intent). We hit this running a local model first and falling back to a cloud one on tool-heavy agent runs: the fallback silently produced a plausible-looking response with no tool call at all, because it interpreted the same instruction as answerable from text. No error, just a wrong answer that passed review because it read fine. Worth treating as a distinct risk from context loss, since it doesn't show up until the tool call was actually needed.