Shield-Consistent Double Dueling Q-Learning with Auditable Inventory Risk Budgets for Trade Execution

Large-order execution agents often propose actions that are subsequently changed by a risk-control layer. This creates a training-deployment mismatch: the value network estimates outcomes for unfiltered orders while actual execution follows the shield. We study shield-consistent Double Dueling Q-learning (SC-DDQN), which trains on the transitions generated by the same observable execution shield used at deployment, and extend it with an adaptive inventory-exposure envelope (BSC-DDQN). The shield operationalizes flow-sensitive urgency, schedule recovery, and deadline floors; the new envelope tightens the permissible post-trade inventory as observed volatility rises. We implement a finite-horizon, six-action execution environment with concave temporary impact and four persistent latent liquidity regimes. Three independently trained seeds are evaluated on 720 paired held-out synthetic paths against deterministic schedules, unshielded DDQN, post-hoc shielding, and shield-consistent ablations. SC-DDQN attains 15.05 basis points (bps) mean implementation shortfall, compared with 18.67 bps for time-weighted execution and 15.88 bps for post-hoc shielded DDQN; the risk-budget extension instead reaches 16.24 bps but reduces inventory exposure. A seed-level sensitivity analysis yields a 95% interval of [-0.50, 2.16] bps for the 0.83-bps advantage over post-hoc shielding; only three independent fits are available. We report tail losses, intervention frequency, and impact-coefficient sensitivity. All outcomes are generated by a disclosed simulator; they measure the interaction between learning and explicit execution constraints rather than realized venue profitability.