fix(tasks): a requeued task starts over instead of resuming the run you cancelled

[Adam-Dalloul]

Fixes #649: start a task, cancel it, requeue it, start it again, and it spins for a while and then lands in 已完成 as canceled. The same task never runs again; only a brand new task works.

Cause: requeue_canceled moves the row back to todo and clears the plan, the failure reason and the pending merge, but keeps conversation_id. launch_mode_for reads exactly that column to choose between Retry and Fresh, so the next start is dispatched as a resume of the run the user just cancelled: it loads the killed agent session, and it reuses the conversation row whose status is still cancelled. The 30 second reconcile sweep reads that stale status as this generation's verdict and settles the task canceled, which is the 已完成 column. Nothing ever clears conversation_id, which is why a new task is the only way out.

The change: a requeue drops the conversation link. The worktree is still reused, and the old conversation still exists in its folder; only the session pointer goes, so the next start runs the task from the top. retry (failed to queued), which deliberately continues the same session, is untouched, and the test asserts that.

The first commit is the failing test on its own, so the CI run on it shows the defect: a_requeue_drops_the_canceled_run_s_conversation reads Some(41) where it expects None.

One consequence worth calling out: while a requeued task waits in todo, the "View session" footer action is hidden, because there is no bound conversation until the next run starts. That matches what the task now is, a fresh unit of work on the board, but it is a visible change.

[Adam-Dalloul]

The CI run on the first commit (test only) is the reproduction: a_requeue_drops_the_canceled_run_s_conversation failed with "a requeued task must start over, not resume the canceled session", 3574 passed / 1 failed. The second commit is the fix.