When Q4 Quietly Breaks a Task

A team quantizes their self-hosted coding assistant model to Q4 to fit it on available hardware. General chat quality looks fine in manual spot checks, but a few weeks after rollout, users report that generated code "looks right but doesn't compile" more often than it used to on the same model at full precision.

1. Explain why Q4 quantization specifically could be the cause here, connecting it to what quantization actually costs.
2. Why might "manual spot checks looked fine" have missed this before rollout?
3. What would you do to confirm the diagnosis and decide on a fix, without simply reverting to full precision everywhere?