Previous audit allowed up to 5 requests in a 30s window — generous,
hid real regressions. Tighten:
- 5s settle after each scene becomes interactive (covers the burst
of mount-time fetches every scene legitimately fires)
- 55s observation window after settle, sized to fit between two
consecutive 60s topbar badge polls
- MAX_REQ_PER_WINDOW = 0; any /api/* or /internal/* request during
the window fails the assertion
This is the operator-experience contract: when you stop typing,
canvas stops fetching.
canvas-bot