GLM-5 stress-tests Astro Builder prompt

This is the right kind of demo for GLM-5: not a polished benchmark slide, but a messy prompt that shows whether the model can hold structure, constraints, and sequencing together.

• –Z.ai says GLM-5 is built for complex system engineering and long-range agent tasks, and its docs claim top open-model results on BrowseComp, MCP-Atlas, and τ²-Bench.
• –The migration guide explicitly tells teams to regression-test randomness, latency, and tool-call streaming, which is a quiet admission that production reliability still needs engineering discipline.
• –Builder prompts are a brutal proxy for real work because they force the model to manage layout, state, and multi-step dependencies instead of just generating fluent text.
• –If GLM-5 handles this cleanly, that is a meaningful signal for scaffold generation and refactoring workflows; if it slips, the gap between benchmark strength and production trust stays real.