Turns out I didn't need to look for exotic formats to solve this problem.
I had physical copies of the disks with 10 sectors per track. Since I couldn't get tar to line up, I posited that there must be missing sectors. Turns out I was missing something, but it wasn't sectors. That turned out to be a dry well.
What's going on is actually a lot simpler. I managed to get Venix installed on my Rainbow (I'll write up how in my next blog, it's fun to login to my rainbow from upstairs while it's down in the basement).
On the Rainbow 100, under both CP/M and MS-DOS there's an interleave of 2, but no per-track offset. This is done by having the logical sectors ride on top of the physical. So, if I write logical sectors in the order 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, it gets put on the disk in sectors with the labels 1, 6, 2, 7, 3, 8, 4, 9, 5, 10. Why go through this crazy arrangement? For performance. This helps sequential performance. After I read sector 1, I have half a disk rotation (at 300rpm or 5/s that's about 10ms) to queue up the read for sector 2. If I have a slow system, this can help performance a lot, since I don't have to wait for the disk to rotate a full rotation if I'm not fast enough to setup the transfer for second immediately. This is often impossible on slower systems since the inter sector gap is maybe 20-40 bytes. 40 bytes transfers at 250kHz, 320bits at 250kHz is about 100 micro seconds. At 4.8MHz that the Rainbow runs at, this is 5000 instructions. But since there's latency between when the transfer is complete to the controller, and the system interrupts this can easily have less than 10 microseconds to respond. On the Rainbow, it takes tens of microseconds to setup the command to read the floppy. Given such a tight time budget, you can get a big gain by requiring only 100ms to get the next sector instead of 200ms.
This is well known and well documented in the Rainbow community. I had to cope with it when I wrote IMPDRIVE back years ago by translating the requested sector so that it would read the proper sector when reading 3.5" floppies which don't do this.
There's another way to get better performance. MS-DOS and CP/M don't do this. But Venix does. If you want to reduce the amount of time that you have to wait for sector 1 to spin by the head when stepping from the previous track, you can skew the sectors another way. If you shift the start of the interleave sequence above to be something like 8, 4, 9, 5, 10, 1, 6, 2, 7, 3. This gives you half a rotation to get the head moved and settled. If sector 1 is right under the head by the time the head is done moving and the host gets its command to the floppy controller, we don't have to wait an average of half a rotation.
So to sort out this, I dd'd one of the disks under Venix ono the disk. Tar was able to read it readily there, but couldn't on raw images. So, I wrote a program to undo these effects. Once I did that, I was able to read the images with tar.
Once I was able to decode the images, I discovered that the floppies I'd read out under dos using rbimg. icreate.exe is used to read it out, and iwrite.exe can be used to recreate them. These days, this is the gold standard to preserve disks. Well, non-copy protected disks. If you want to preserve the disk in all its glory, use a disk imaging solution like Kryoflux.