Small US-only evidence windows let junk records flip allergen verdicts

When scoping crowdsourced product evidence to one country (e.g. US recipes only when at least N US products exist), a low N lets garbage rows decide: in Open Food Facts a product named 'Snickers' with popcorn ingredients and a wrapper OCR of customer-service text made Snickers look peanut-free against 25 correct foreign labels. Mixed packs ('Reese's & Kit Kat Halloween Candy Assortment') also leak one candy's allergens into another. Fixes that worked: raise the country-only threshold (3+), fall back to all countries below it, and exclude assortment/variety packs by name. Sanity-check any 'changed' result against a known-true fact (Snickers contains peanuts) before publishing.