On further debugging I am forced to conclude that the Ogg container format is lame. This is not saying anything about the Vorbis compression format; I haven't delved into that other than to note that it has quite a bit of overhead for small files, at least the way Audacity encodes it.
Here is a good explanation of the ways in which the container format is unfortunately designed. For my purposes, the main problem has to do with how packet size is encoded. They didn't want to put a fixed-size 32-bit value in the packet header because they expected typical packet sizes to be under 256 bytes. Never mind that they have at least three other wasted bytes in their header, and they chose an encoding scheme that made their header variably-sized. The encoding scheme for packet length is this (feel free to laugh): first, a count byte. Then an array of that many bytes. Each byte represents a part of the size count; to get the full size, add them all together. You are expected to set the first N-1 bytes to 255 and the last byte to the remainder. It's sort of the inverse of compression, really. They should have thought very, very hard before making their packet headers variable size. And then if they really, really thought it was a good idea they could have used the UTF-8 style of encoding arbitrary-size numbers.
The decoder I was reading followed the spec in that it assumed that a non-255 value indicated the end of the array (it isn't paying attention to the count as far as I can tell). However, all the files I've encoded with Audacity have an early packet with a non-255 value in the initial byte, followed by a bunch of 255s, followed by a non-255 final byte. The initial non-255 byte is causing the decoder to prematurely think it's reached the end of the packet.
It's telling that Wwise (an audio middleware provider with which I am familiar) supports Vorbis compression but wraps it in a RIFF-style container format (what .wav files use) instead of using the Ogg container format. (Confusingly they retain the .ogg file extension, though.) It'd be nice if there was a standard for this; I'll have to investigate.