Attention from first principles: queries, keys, values and the causal mask

Derive self-attention from the idea of soft lookup, implement causal scaled dot-product attention in plain Python, verify that future tokens cannot influence the past, extend to multiple heads, and count the cost.