Alpha-v2.4 is an N-agnostic refactor of the v2.x stack. The core idea: branch (per-symbol decision module) and V_global (portfolio observer) become architecturally independent units. The branch is trained in isolation by its per-symbol Q signal; V_global reads the branch's rich features and produces a portfolio-level state-value estimate, with no gradient feedback into the branch.
Why. v2.3 fixed the per-sym vs global Q conflict by detaching qg_new in the actor loss, but the global head still lived inside the shared critic encoder — polluting per-symbol features with portfolio-level signal. v2.4 lifts the global value out into a separate observer module, making each branch a self-contained, composable unit.