If whatever adapter you're using isn't working in MASTER, there's a bigger problem, or the adapter is damaged.

I've also experienced what you are saying--that when the adapter is jumpered as "Master" it cannot cooperate with a "Slave" hard drive on the same cable.
It's not always apparent whether these things are the adapter's or the card's fault.
A CompactFlash datasheet suggests the cards have CSEL# pin. When left floating, the card operates as "Slave" while when grounded it becomes "Master".

IDE drives sharing the same cable also use the PDIAG# and DASP# pins at powerup to handshake.

Beyond that, "master"/"slave" is in software, when the host flips a bit in the Drive/Head register to switch what drive it is accessing. It should be hard to mess that up, so I'm surprised they detect correctly in your situation (in mine, everything completely stops working when booting with a "slave" connected) but then fail later.

I think you will need to have 2 IDE channels and have the HDD and CF both as master on independent channels. If the CF is Channel 0 master it will take over and be the boot drive if inserted. If the CF is Channel 1 Master it will appear as a second dive and continue booting from the HDD when a CF is present.

I had to do this with my SD/IDE adapter as it can only work in Master.

Hello,
I can confirm:
a "Kingston 2GB 133X CompactFlash Elite Pro" works jumpered as "slave"
a "Transcend CompactFlash 80x 512MB" does not.
It's purely a card issue, not the adapter.

In fact, with a WD 80GB IDE drive as "master,"
the Transcend 512MB CF with an adapter as "slave,"
and an ISA Multi-I/O card,
when an ATA IDENTIFY is performed against the Transcend CF, various bits are flipped in the response (concentrating specifically on the word Transcend, but it's everywhere). Different bits are flipped on each attempt. Replacing the rep insw to have some delays (out 80h,al for example) between each word read seems to reduce the bit flips, but not eliminate them.

Is there some inherent fight between pull-up/pull-down resistors and the opposite drive asserting data lines when two IDE drives are sharing one cable?

I finally found the solution:

Both devices (hard drive and IDEtoCF drive bay) are connected as master on the primary IDE channel. This way, without a CompactFlash card inserted, the main hard drive is recognized as master without issue, and when the CF card is inserted, it becomes master, disabling the hard drive.

It's not the most elegant solution, but it works, and it does exactly what I need. I tested booting from both devices, and it works.

I found some 4GB SanDisk CFs that I forgot I had ordered. They work as both "master" and "slave". So it's the card, not the adapter, that fails to work as "slave".