RoPE Playground — Argand rotation + relative-position score

Toy 2D planes, but with an option to sum multiple frequencies (closer to real RoPE’s “many 2D pairs”).
We assign each token a fixed “content” vector (deterministic hash → angle + magnitude), then apply RoPE rotation by position.
p = index + start
Each plane uses θᵢ = base · 1/10000^(2i/K) (sinusoid-style spacing).
base content vectors RoPE-rotated vectors (for current θ) query token m

Score vs relative offset Δ = n − m

Score is a toy attention-style dot: score(n) = (R(p_m) q_m) · (R(p_n) k_n).
In multi mode: we sum scores across K independent 2D planes (mimics many RoPE pairs).