Three different things get called the same word.
A restart gives you a new shell in the same directory. Whatever was running is dead. A lossy reattach means the session survived and you get its current screen back, but everything that scrolled past while you were away is gone. That is what most implementations do, and it is where the interesting output usually was. A lossless reattach gives you the session and its full history, exactly as if you had been watching the whole time.
Why the third one is the only useful one here
The whole reason to leave an agent running is to read what it did afterwards. If reattaching gives you the last twenty-four lines, then the compile error at line four hundred (the reason you came back) is not there.
What it takes
The pty bridge must keep a bounded buffer of the byte stream, not of rendered frames, because terminal output is a stream of control sequences and replaying it is what reconstitutes the screen. On attach it replays the buffer and then switches to live.
The bound is a real decision: too small and long builds are truncated, too large and a chatty process eats memory. It is a per-session cap, not a global one.
Where it still breaks
If the bridge reconnects but the interface does not reconcile which sessions exist, you get tabs that are present and blank, attached to nothing. It is worth testing explicitly, because it looks like a rendering bug and is a lifecycle one.