This is often a slippery slope. Assumptions tend to erode over time, eg. I've seen code assuming ASCII-only input fall over due to people copy/pasting from Word

This is often a slippery slope. Assumptions tend to erode over time, eg. I've seen code assuming ASCII-only input fall over due to people copy/pasting from Word (which automagically turns hyphens, quotes, etc. into non-ASCII). Also, whenever there's a lot of data being looped over, it can be very tempting to add new features to the existing loop, rather than adding a new pass (inefficient) or trying to collate the results after they've been spread out (complicated & error-prone). This can turn a simple parser into a core piece of business logic, and of course it's then only a matter of time before custom, XML/JSON-incompatible 'directives' start creeping in to control that business logic.