queen(actors): seeded deterministic simulation of the actor runtime as a CI gate #1721

Part of #1712 (item 5). Spec: gHashTag/t27#8279 (specs/queen/simulation.t27).

Problem

The benchmark that cleared the actor reviewer (#1698) ran on a virtual clock and a seed, and still missed the go-live hot loop: it modelled no unreadable row, so about 184 visits a minute against 22 unreadable rows were found only in production (gHashTag/t27#7851). Nothing searches the actor runtime's interleavings and faults, and nothing checks that a seeded run is reproducible.

What

A seeded deterministic simulation of the real actor runtime, run as a CI gate:

• a harness under trios/agent-server/apps/server/tests/sim that drives createActorSystem, the supervisor, remoteChild, createMemoryNet and the reviewer actors on the virtual clock, with every random number, each step's event and the fault mix taken from the simulation card;
• invariants checked after every event: no pid reused while mail to it is on the wire, no turn for an exited process, supervisor intensity respected, review concurrency within its cap, every restart after the previous incarnation ended, no hot loop on a row;
• every seed run twice, and the two logs compared step by step;
• a failure prints the seed, the step, the minimal log tail and the command that replays it;
• a CI job for PRs into actors-next that touch the actor runtime, and the same group in Tests for queen.

Acceptance

• bun test tests/sim passes on actors-next with this change.
• Removing the wait backoff (the go-live defect) fails the gate from a seed.
• The PgLink defects of Queen actors: seven improvements from the competitor study (integration branch actors-next) #1712 item 6 reproduce from a seed, before they are fixed.
• Time per seed and in total is reported.