@effect/cluster: Shard acquisition loop never completes with BunClusterSocket SQL storage

[Necmttn]

Bug Report

Versions

• effect: 3.19.15
• @effect/cluster: 0.56.1
• @effect/platform-bun: 0.87.1
• @effect/sql-pg: 0.50.1
• Bun: 1.3.10
• PostgreSQL: 17

Description

When using BunClusterSocket.layer({ storage: "sql" }), the runner registers in PostgreSQL and receives shard assignments, but never transitions to entity/cron registration. The shard acquisition loop runs indefinitely.

Expected behavior

Runner starts, acquires shards, starts ClusterCron entities, logs "tick" every 10s.

Actual behavior

Runner logs Shard acquisition loop / New shard assignments / RunnerStorage sync forever. Entities never register. The runner is alive but non-functional.

Reproduction

git clone https://github.com/noktadev/effect-cluster-shard-stall-repro
cd effect-cluster-shard-stall-repro
bun install
bun run db:up
bun run start
# Wait 30+ seconds - no "tick" output appears

Repo: https://github.com/noktadev/effect-cluster-shard-stall-repro

Minimal code

import { ClusterCron, ClusterWorkflowEngine } from "@effect/cluster";
import { BunClusterSocket, BunRuntime } from "@effect/platform-bun";
import { PgClient } from "@effect/sql-pg";
import { Cron, Effect, Either, Layer, Redacted } from "effect";

const TickCron = ClusterCron.make({
    name: "TickCron",
    cron: Cron.parse("*/10 * * * * *").pipe(Either.getOrThrow),
    execute: Effect.log("tick"),
});

const PgLive = PgClient.layer({
    url: Redacted.make("postgres://postgres:postgres@localhost:25432/cluster_test"),
});

const RunnerLive = BunClusterSocket.layer({ storage: "sql" });

const EnvLayer = Layer.mergeAll(TickCron, ClusterWorkflowEngine.layer).pipe(
    Layer.provideMerge(RunnerLive),
    Layer.provideMerge(PgLive),
);

Layer.launch(EnvLayer).pipe(BunRuntime.runMain);

Observations

• PostgreSQL cluster_runners table shows the runner registered with healthy: true
• Shard assignments ARE received (visible in DEBUG logs as New shard assignments with shard IDs)
• The runner never progresses past shard acquisition to entity registration
• SingleRunner.layer({ runnerStorage: "memory" }) works perfectly
• Truncating cluster_runners/messages/replies/locks and restarting sometimes (but not always) resolves the stall
• Ghost runners accumulate across restarts (no deregistration on pod/process shutdown)
• Discovered in production Kubernetes with Bun-based runners