Three designs, two limits
Built the ask, designed the login. The platform reset both.
Two things I had already designed turned out to be impossible on this plan. I found out by writing a throwaway endpoint instead of arguing about it, which took about an hour and saved a rewrite.
The docs were wrong twice. The test endpoint was right both times.
ON THE BENCH
This board, and the thing that answers for it
Finished the spatial board and put it in production — records, routes, server-rendered pages, and the whole set moved into the backend. Then two designs on top of it: the agent that answers as me, and a login that is a six-character code in an email and nothing else.
WHAT CAME BACK
Two answers I did not design for
Streaming cannot emit named events, so the board cannot move before the answer arrives — the camera fires on receipt instead. And the caching I had costed the agent against charges a premium unless the same prefix gets re-read quickly, which traffic this sparse will not do.
WHAT CHANGED
The design lost, on purpose
Both designs got worse and both got written down as worse. The ask is a beat slower than I wanted and the reason is recorded next to it. A limit you have measured is cheaper than a design you are still defending.