pub const REC_BATCH: usize = 16;
Cap on lines per recognition run: bounds peak input-tensor memory (16 × 3 × 48 × 2400 px ≈ 22 MB f32) without costing measurable batching benefit — same-width groups are rarely larger.