> why not just distribute your message over that channel in the first place

Latency? You can hand deliver a password ahead of time, but not messages.

One-time pad isn't a password. It is a flash drive or hard drive full of random bits.

The difference between those is just one of scale and storage.

You still have to reliably move a chunk of out-of-band information in a way such that it gets to (and only gets to) the person you want to have it.

The difference between one-time pad and stream cipher is provable, absolute secrecy, and really good secrecy. If don't care about that, there is zero point to one-time pad.

Also, it isn't just a "chunk", for one-time pad it has to be the same length as the messages. Which is fine if just short messages but a lot harder if lots of data.

If can exchange lots of data, better off using them as keys for stream cipher.

One-time pads are obviously not a serious widespread cryptography proposal.

But the question of, "Why not just send the message instead of the pad" is pretty straightforward: when you have the opportunity to safely deliver the pad, you don't know what the message will be. When you do know what the message will be, you don't have the opportunity to safely deliver the pad.

Because with QKD you can distribute a random key knowing that there were no observers but you cannot distribute a message with the same guarantees. Specifically, any given bit exchanged might be observed, but that is detectable so the bit can be discarded.

I read some years ago about a non quantum technique to achieve the same based on (I think) noise in a coupled electronic system. I wonder if that has been tested further.