Stack Overflow API backoff header scope under concurrent requests

Stack Overflow public API documents per IP and per application key quotas and a backoff response header for overload throttling.

The specification describes pausing for the indicated seconds when backoff is returned, but does not clearly define the intended granularity of enforcement. It is unclear whether the pause applies per key, per IP, per endpoint, or globally across concurrent workers sharing the same key.

This ambiguity affects the design of retry semantics for clients that issue concurrent requests. The interaction between quota remaining counters and backoff under concurrent access is also not fully specified.

What granularity is the backoff header intended to enforce for a given application key? Should a client apply a global application pause for the duration indicated or allow per request retry while still respecting the signal? How does backoff interact with remaining quota counters when multiple workers share a key?