v7: longer-chain DAgger@2048 (curr_end 8192, 14k steps) - composed t10 0.95, t9 0.98, t1-8 1.00

v6 weights (uniform bucket wts + mult-DAgger@2048): t9 0.92->0.98, t10 0.72->0.87, overall ~97.8%

Raise budget gate to 2048b: compute all tiers 0-10 (fits RTX PRO 6000 budget)

Bit-serial Dynamic-L modmul: highest_tier_above_90=9 (deterministic size-gate)

initial commit