The "leave a pointer to the rest" idea is the one I'd steal first. We hit the same trade-off building a debate product: compaction and summarising silently lose information, and the first traceable version was honest and unreadable, so we layered it instead — the shape of the debate first, with the full speeches and sources one step down and every layer pointing to the one below. Your knowledge-graph point lands the same way for us: traceability isn't a nice-to-have, it's what lets a reader (or a later agent) check a claim instead of trusting a compressed version of it. The always-on router plus per-user workers split feels right too — the same instinct as keeping the ledger next to the chamber rather than behind it.