• Secondary rate limit (is_secondary_rate_limit: a 403 or 429 with retry-after or a message naming the secondary rate limit, which a GraphQL 200 also carries): waits retry-after (seconds or an HTTP date), or, when it is absent or not positive, 60 s doubled for each retry already made ("wait for at least one minute", "an exponentially increasing amount of time"). The policy's Retry-After cap does not apply; its retry count and the call's remaining time (RetryContext::remaining_ms, stopping with Budget) do.
• Primary rate limit (a 403 or 429 with x-ratelimit-remaining: 0, or a GraphQL RATE_LIMITED error): waits until one second past x-ratelimit-reset ("you should not retry your request until after" it), and only once per call, as octokit does; a second one stops with Exhausted. Without a readable reset it stops with NotRetryable rather than guess, and the policy's retry count and the remaining time bound it as above.
• Transient GraphQL failure ("Something went wrong while executing your query", raised as Decode by graphql for a document that is not a mutation, in the "graphql" bucket): waits the policy's backoff, within its retry count and the remaining time. A mutation's ("graphql:mutation") is never retried, since it may have been applied.

test {
  let retry : @github.GitHubRetry = {
    policy: @runtime.RetryPolicy::conservative().with_max_retries(3),
  }
  let post = @http.Request::post("/repos/o/r/issues")
  let limited = @runtime.Status(
    status=403,
    headers=@http.Headers::new(),
    body=b"{\"message\":\"You have exceeded a secondary rate limit.\"}",
  )
  let waits = [0, 1, 2, 3].map(attempt => {
    retry.decide(limited, @runtime.RetryContext::new(post, attempt~))
  })
  assert_eq(waits, [Wait(60000), Wait(120000), Wait(240000), Stop(Exhausted)])
}