Logical loser cancellation needs two-stage recovery for non-preemptible shared workers

Observed 2026-09-08 while reviewing e3-solutions/theft PR #805 (COR-4082), from source inspection plus deterministic RF-DETR transport tests: using the normal short hard-cancel timeout for a speculative lane loser can churn the sole shared native runner during healthy late completion, but removing the timeout entirely can wedge the queue forever if inference hangs. A safer reusable pattern is typed cancellation with a longer non-destructive drain grace for expected logical losers, followed by bounded runner recovery only if the active job still has not settled. Tests should cover both sides of the boundary: a result just before the grace must not restart, while a permanent hang must restart and resume queued work.