OK - so a few points based on the comments:
1.) By definition Ethernet is not routed but rather switched, bridged or repeated.
2.) As mentioned, protocol reliability implies the ability to retransmit lost and errored frames without the interaction of a higher-layer protocol. This means that the sender and receiver will need to maintain some kind of state to track packet sequence and likely some amount of flow control to adjust based on buffering. This also means you'd need semantics for session establishment and teardown - which means timers. This is much heavier in actual practice than UDP/IP. It's also worth noting that there are lightweight/embedded IP+UDP stacks that would prevent you from having to reinvent the wheel.
3.) Take a look at the RoCE family of protocols - super low latency and supports your requirements. Interestingly version 1 of the standard runs over raw Ethernet and supports jumbo frames. Whether it hits your requirements for reliability is another question. IMO the more interesting point is that ROCEv2 has similarly excellent latency characteristics but runs over UDP (and is well-supported by a number of NIC and switch vendors). It's worth noting that part of the operation of the RoCE protocols requires lossless Ethernet facilities. Minimally this includes PFC but the spec also calls out some of the other DCB specs.