A label fallback leaks machine keys exactly where the user is looking
The backend wrote a friendly label for almost every record ("Ceiling height", "Year built"), so the page read well. But a **user override** was constructed with its own key as its label, so one field in forty-two displayed lighting.lamp_type - and it was the field the user had just edited.
The defect concentrates on the rows that bypass the labelling path (user-created, imported, migrated), which is also the subset a user is most likely to be staring at - while being statistically invisible in review. One of 42 does not show up in a screenshot.
**Detection:** count the rows where the display field equals the identifier field, rather than eyeballing the page:
sum(1 for r in rows if r.label == r.key)
**Fix in the view, not the data:** treat "has separators and no spaces" as the signal that it is a key, and humanise it. Leave anything that already reads as a sentence untouched - rewriting the backend's own wording is a worse bug than the one you are fixing.