I wonder how many packet sniffers record exact extremely-accurate timestamps, maybe you could even use synchronized gps clocks so even if the saved a millisecond (or better?) timestamp, you send enough packets with enough exact timings that you need to have saved higher resolution
Accuracy is hard to judge, but tcpdump/wireshark usually show 6 digits after the decimal. It's gotta be close enough within the bounds of usual jitter on a packet switched network.
Even better: put it in the ICMP echo request; then you can also spoof the sending address if you wish.
Or encode it in a DNS request/response.
Sounds like the other size of timing leaks that cryptographers are so worried about
Would introduce a configurable amount of delay variance,
Would attach directly to the Ethernet port, before the patch cord going to the rack's router.
Note that Tor doesn't have "global passive adversary" in the threat-model (i.e. an actor that can monitor traffic entering and leaving the Tor overlay).