The fragment identifier is made of a hash of the contents of the text node in the DOM. For instance, imagine the following html:
<span>Example text, with some <strong>bold</strong> and <em>italics</em>!</span>If the user selected starting at the word text and ending at the word italics (but not including the exclamation mark), the fragment identifier would contain two hashes: one of the string Example text with some , and the other of italics. It would also contain two offsets: 8 for the index of the start of the selection (text starts at position 8 in the string Example text, with some ), and 7 for the end of the selection in the sting italics. There is also a version number, which is used so that the exact format can change as needed, while still supporting old links.
This has some obvious drawbacks — if you have very large paragraphs, changing any text within them will break all links to that paragraph. I have some ideas for reasonable solutions to this, but haven't started implementing them yet.
Handling of nodes with the same text contents (and thus the same hash) is described here.
The exact details of the hashing and format are unimportant, but are commented in the code for the curious.