Memoize with a Time-to-Live

Calling an LLM or API with the same input twice is wasteful, but caching forever serves stale results. A TTL cache remembers results for a limited time.

Implement a decorator factory ttl_cache(seconds, clock=time.monotonic):

• Cache results keyed by the call's positional and keyword arguments (keyword order shouldn't matter).
• A cached value is reused while clock() - stored_at < seconds. After that, call the function again and refresh the entry.
• Exceptions aren't cached.
• The wrapped function gets cache_clear() (empties the cache) and cache_info(), returning a dict {"hits": h, "misses": m, "size": n}.
• Preserve the function's name with functools.wraps.