custody-check: stop-line misses health and children's signals in values and prompt text #14

Problem: the stop-line ("health data, children's data, payments → get a person") only fires from the founder's own tick or from field names like medical. Two real apps showed the gap:

• a health-adjacent app whose signup form offers health-related options as values in a dropdown;
• an AI app for children whose system prompt text names a child's age and diagnosis.

Neither raised anything, while both belong behind the stop-line.

Why it matters: the stop-line is the one custody-check answer about safety, not accuracy, and workshop founders will bring exactly these apps.

Direction (to be researched first):

• Scan option values, labels, copy and prompt text for health, age-of-user and payment signals, not only field names.
• Evidence only: this raises a stop-line hint for the founder to confirm and never decides the door alone.
• Check it against the established data categories (for example health information, children's data, special-category data) instead of a hand-made word list.

Done when: a fixture with health values in a select, and one with an age and a condition in a prompt string, each produce a stop-line hint; a clean fixture produces none; no false stop-line on the existing corpus.