Open-set semi-supervised learning (OSSL) addresses the practical scenario in which unlabeled data contains out-of-distribution (OOD) samples. Early OSSL methods typ- ically relied on strict hard-filtering, which frequently led to severe over-rejection of valid samples and poor precision. To mitigate this, recent approaches have adopted soft-filtering or explicit outlier modeling; however, these methods often merely shift the failure mode along a pronounced precision-recall trade-off. Specifically, methods that learn a dedicated unknown-class representation overfit to seen OOD patterns, while soft-filtering methods suffer from an ID-dominated optimization where outlier scores only attenuate gradient magnitude. Because this magnitude-only reduction fails to alter the optimization direction, OOD updates are progressively absorbed into known-class regions, ultimately leading to poor recall and under-rejection. To overcome this opti- mization limitation, we propose OwareMatch, which integrates open-set knowledge directly into the classifier’s gradient flow. By decoupling logits into ID and OOD energy components and asymmetrically modulating gradients, OwareMatch progres- sively decorrelates the ID and OOD optimization paths. This gradient-level decoupling suppresses OOD absorption into ID decision boundaries, instead encouraging OOD features toward a class-agnostic rejection region. Because this rejection mechanism avoids explicitly learned, seen OOD-specific representations, it is less distribution- dependent and transfers more robustly to unseen OOD shifts. Extensive experiments demonstrate that OwareMatch effectively mitigates the precision-recall trade-off, main- taining competitive closed-set accuracy while consistently improving OOD F1-score and robustness under both seen and unseen shifts.