Fix the stale-while-revalidate cache policy

Your serverless function fetches a user's feature flags from an API. An AI's first cache was a module-level Map, which starts empty on every cold start, and its 1-hour TTL made every request after expiry hit the API at once. The new design keeps one shared cache entry and serves a stale value while one refresh runs, so callers do not stampede the API.

An assistant wrote plan_cache_reads(request_times, fresh_ttl, stale_window), which decides what each request does. It is in the editor. It has a bug. Find it and fix it.

Input and output:

• request_times: when each request arrives, in seconds, in non-decreasing order.
• fresh_ttl: how long a stored value counts as fresh.
• stale_window: how long after that it may still be served while it refreshes.
• Return one action per request: "fetch", "hit" or "stale_revalidate".

Rules the tests check:

• A fresh_ttl or stale_window below 0 counts as 0.
• The first request is a "fetch", and the entry is stored at its time.
• For every later request, the age is its time minus the time the entry was stored:
  • age below fresh_ttl: "hit".
  • otherwise, age below fresh_ttl + stale_window: "stale_revalidate". The stale value is served and one refresh starts, and from now on the entry counts as stored at this request's time.
  • otherwise: "fetch", and the entry is stored at this request's time.