recall-first rebalance: realistic composition (1-2 relevant + content-heavy), category labels

v3-labels: category-match relevance (validated 11/11 vs human); corrects over-strict field-overlap labels

v2.3: English-text cleanup (fix lang filter), strip junk fields + title cruft, drop empty titles (361 sessions)

v2.2: add dense-positive slice (5-8 relevant tabs) for the recall@m sweep

v2.1: extend n-axis to 50 for size diversity

v2: provides-based relevance + content-source tabs + n-axis (456 sessions)

Add smart-form-fill relevant-tabs (select_tabs) eval dataset: 138 sessions