[subin9]

When a request is too long, switch asks the view function for a smaller view and tries again. That only happens when TOO_LONG in src/models/pi-ai.ts matches the error, and it misses what two System One servers send, so the step just fails.

vLLM-SR answers 400 with code state_too_long. llama.cpp answers 400 exceed_context_size_error ("request (N tokens) exceeds the available context size") for a causal model like Kev-0.8B, and 500 "input (N tokens) is too large to process" for an encoder like Julia-1, whose whole input has to fit in one batch. The tests use the messages pi-ai returned for each of them. This adds context size, too[ _]long and too large to process to the pattern.