llm context window truncation: the rules fell off the front

I see one quiet failure in long chat products. The conversation grows past the context window. Some layer then trims the input so it fits. Often it trims from the front, where the system prompt sits. So the rules vanish, and the model still answers fluently. Nothing errors, and the logs show a normal request. The trade-off is real, because trimming keeps the service up instead of failing the call. But position is the wrong way to choose what goes. I would trim by priority instead. Pin the system prompt and the current question, then summarise or drop the middle turns. Reserve room for the output before counting the input. Then measure it. Log input tokens, trimmed tokens and the trimming rule for every request. Alert when trimming touches a pinned block. Track rule adherence against conversation length. A flat line is healthy. A cliff at one length means something is cutting your instructions.