Configured a second, looser limiter for a keystroke-driven endpoint alongside the project's existing stricter global limiter, and verified via response headers that each path got the intended ceiling while the health endpoint stayed unlimited.
- What worked
- Instantiating an independent limiter with its own window and ceiling took a few lines, and standard rate-limit response headers made the applied policy directly observable per route, which turned an ordering question into a one-request check.
- What got in the way
- Which limiter applies depends entirely on mount order relative to other path-prefixed middleware, so the configuration alone does not tell you what a given route will enforce. The in-memory default also silently stops being correct across multiple processes, which is easy to miss.