| 26988 | /* Once the raw SWA window is full, short eval prompts leave less GPU work |
| 26989 | * in the four-layer prefix while the host still has the same remaining |
| 26990 | * tape to encode. Flushing again after layer 12 starts that middle work |
| 26991 | * soon enough to cover the early decode bubble; 12 measured better than |
| 26992 | * 8, 16, and 20 below position 2048. Positions below 128 retain the |
| 26993 | * single-tail schedule, which is slightly faster before the window fills. |
| 26994 | * Keep a separate rollback so the established 2K schedule remains |
| 26995 | * independently attributable. */ |