Faithful Context Compression for Agentic Sessions: A Reproducible Benchmark and a Granularity-vs-Signal Decomposition

Context compression is widely used to cut the token cost of long LLM sessions, but most evaluation measures only downstream task accuracy and studies single-shot prompts. We ask a narrower, more diagnostic question in the multi-turn agentic regime: when a compressor shrinks a session, do the specific facts a later turn depends on survive? We release a reproducible, judge-free benchmark that plants facts with known ground-truth answers in synthetic agentic sessions and measures token savings against deterministic fact retention, decoupled from any model (and thus immune to pretraining contamination). Using it, we run a controlled comparison (n=900, 3 seeds) of message-level relevance selection (query-aware, training-free) against token/sentence-level self-information pruning (Selective Context) at matched savings. Selective Context is perfectly lossless (100% retention, zero between-seed variance) across the entire savings axis up to 89%; message-level selection is far worse (56% aggregate). Rather than stop at this negative result, we decompose the gap. Holding the selection signal fixed and moving from message- to sentence-level granularity recovers +63.6 points of paraphrase retention (12.5% -> 76.1%); switching the signal from lexical to embedding recovers a further +20.4 points (76.1% -> 96.5%). The counterintuitive finding is that the winning method is query-agnostic: query-awareness is not the differentiator - granularity and semantic matching are, and self-information pruning is near-optimal precisely because answer-bearing spans in this task are high-surprisal. We characterize exactly when a retention metric and a downstream-accuracy metric agree, pre-empt the fact-planting circularity objection, and release the corpus generator, harness, and a public leaderboard so others can reproduce and extend the frontier. Code: https://github.com/nac7/ctx-gate. Benchmark dataset: https://huggingface.co/datasets/nac7/tool-output-faithfulness (CC-BY-4.0, DOI 10.57967/hf/9671).