test(opencode): add regression test for lost terminal-iteration wakeup

[scotttesler]

Your fix in

• fix(runner): queue work when already running instead of discarding anomalyco/opencode#36375

is the right root-cause repair: adding the RunningThenRun state so ensureRunning queues new work instead of discarding it while the runner is already Running is exactly what stops a background-subagent notification from being lost. This PR adds only one regression test on top of your branch — no production changes.

Your tests already cover a single queued run: queuing behind a running turn, the RunningThenRun transition, and a third caller sharing the pending run. What none of them exercise is a repeated re-queue — a second notification that arrives after the first queued run has already been promoted to Running, forcing the runner to cycle Running → RunningThenRun → Running → RunningThenRun → Running. That's the real-world case of a parent session getting several background-subagent wakeups over its lifetime, so it's worth locking down.

The new test re-queues each run that arrives after the previous queued run is promoted (in packages/opencode/test/effect/runner.test.ts):

• starts a parent turn and waits for Running;
• queues notification A while the parent is busy and waits for RunningThenRun;
• releases the parent and waits for A to be promoted back to Running;
• queues notification B while A is running and waits for RunningThenRun a second time;
• releases A, then asserts each caller receives its own result (parent, notify-a, notify-b), the work ran in order, and — via a flag A sets on completion that B reads at its start — that B was promoted only after A completed (serialization, not just start-ordering), ending back at Idle.

It synchronizes only on published state via waitForState, with no fixed sleeps, per the package's test-synchronization guidance.

The test fails on the pre-fix runner and passes with your fix. With the fix present:

$ bun test test/effect/runner.test.ts -t "re-queues each run that arrives after the previous queued run is promoted"
1 pass
0 fail
5 expect() calls

The full runner.test.ts file is 29 pass, 0 fail, and tsgo --noEmit, prettier --check, and oxlint are clean.

With only packages/opencode/src/effect/runner.ts reverted to its pre-fix parent (the test left unchanged):

(fail) Runner > re-queues each run that arrives after the previous queued run is promoted [1006ms]
TimeoutError: TimeoutError
0 pass
1 fail

Pre-fix the runner never reaches RunningThenRun because the queued work is dropped, so waitForState(runner, "RunningThenRun") times out — confirming the test guards the behavior your fix restores rather than passing vacuously.