Retry Decorator with Exponential Backoff

LLM and web APIs fail transiently with timeouts, rate limits and 503s. The standard fix is to retry with exponential backoff: wait 1s, then 2s, then 4s…

Implement a decorator factory retry(times=3, exceptions=(Exception,), base_delay=1.0, sleep=time.sleep):

• Call the wrapped function up to times times in total.
• If it raises one of exceptions, call sleep(base_delay * 2 ** attempt) (attempt = 0, 1, 2…) and try again.
• If the last attempt fails, re-raise that exception, without sleeping after it.
• Exceptions not in exceptions propagate immediately, with no retry.
• Return the function's result as soon as a call succeeds.
• Preserve the wrapped function's name and docstring (functools.wraps).

sleep is a parameter so tests can pass a fake one instead of actually waiting.

Example

@retry(times=3, exceptions=(TimeoutError,), base_delay=0.5)
def call_api(): ...
# on two timeouts then success: sleeps 0.5, then 1.0, then returns