v2 (envoy) actors ignore crash_policy: a lost runner parks the actor (Sleep) instead of rescheduling it #5908

Environment

• rivetkit 2.3.27, rivetdev/engine:2.3.27 (self-hosted, Docker), Postgres backend
• Node 22, one runner pool (default), actors created through rivetkit and through the engine REST API

Expected

versions.mdx says that on SIGKILL "actors are rescheduled with a crash backoff penalty", and
actor-rescheduling.md maps a lost runner to reschedule_actor. Creating an actor with
crash_policy: "restart" should make it restart on another (or the restarted) runner after the
runner is lost.

Actual

After kill -9 of the runner process, the actor's last stop reports crash_policy: "sleep",
error: {envoy_connection_lost}, and the actor stays parked until something sends it a message or
an alarm fires. A workflow in progress does not resume on its own.

• rivetkit's client always sends "sleep" and does not expose the option.
• Creating the actor via the engine REST API with crash_policy: "restart" still reports "sleep".
• Engine source (2.3.27 and main): pegboard/src/workflows/actor2/runtime.rs handle_stopped
  maps StoppedVariant::Lost => Decision::Sleep and only reallocates if an alarm is due;
  pegboard/src/ops/actor/util.rs reports v2 actors as CrashPolicy::Sleep unconditionally.

Related: #3725 (closed, not planned), PR #4822 "restore crash_policy plumbing" (closed unmerged),
#5894 (open; wake ignored while Reallocating — different trigger).

Repro

// runner.ts — start with `tsx runner.ts`, then `kill -9` it while `work` is mid-loop, restart it.
import { actor, setup } from "rivetkit";
import { workflow } from "rivetkit/workflow";

const job = actor({
  run: workflow(async (ctx) => {
    await ctx.queue.next("go");
    for (let i = 0; i < 20; i++) {
      await ctx.step(`page-${i}`, async () => {
        await new Promise((r) => setTimeout(r, 500));
        console.log("page", i, "done");
      });
    }
  }),
  actions: { go: async (c) => c.queue.send("go", {}) },
});
const registry = setup({ use: { job }, endpoint: "http://127.0.0.1:6420", token: "<admin-token>", namespace: "default" });
await registry.startAndWait();

1. Start the engine (2.3.27) and runner.ts; call client.job.getOrCreate(["a"]).go().
2. After ~4 pages, kill -9 the runner; restart it 3 s later.
3. Observe: no further page N done lines. The actor stays asleep (GET /actors/{id} shows
   crash_policy: "sleep", envoy_connection_lost) until it is messaged.
4. Same result when the actor is created with POST /actors and "crash_policy": "restart".

Measured

• Stock 2.3.27, 3 s runner downtime: the workflow did not finish within 300 s (2 of 2 runs, with
  and without crash_policy: "restart" on the REST create).
• Workaround we ship: a durable alarm (c.run.setWakeAt, 3 s, refreshed every 1 s, cleared in
  onSleep) plus pegboard.envoy_lost_threshold = 6 s (default 15 s). Resume after restart:
  3.4–8.5 s over 4 spike runs; 2.6–2.7 s in our integration suite (3 runs); 0 completed steps re-run.
• SIGTERM drain (rolling deploy) is fine without any workaround: resume 0.4–0.7 s.

Question / suggested fix

Is Lost => Sleep for v2 actors intended? If yes, please fix versions.mdx /
actor-rescheduling.md. If not, honour crash_policy: "restart" for v2 actors (and expose it in
rivetkit actor options) so users don't need an alarm-based watchdog to survive a runner crash.