When ORES hits the wall, it starts returning "server overloaded" errors. We have a mechanism for "revoking" tasks, e.g. when the scoring of a particular revision takes too long and we want to avoid potential DoS, however we're abusing it by marking requests made during a general overcapacity interval as revoked. This is incorrect, and makes the requested revisions unscoreable for 24 hours.
We should distinguish between general overload and an expensive request. The former should be refused with a 5xx HTTP status code (currently returns 200 but the contents are all errors), and the revisions in question should not be penalized.