queen actors: where the per-message cost went after the #1712 lanes, and five load-dependent tests #1743

Part of #1729 (night-loop item 8) and of #1712.

The anomaly

On a quiet host the actor runtime's base cost per message, telemetry off, was 4.53 µs without slices and 5.80 µs with them (t27#7851, comment 6090872063). Before the #1712 lanes it was published as 3.7 and 5.5 µs. Where did the ~20% go?

Load-dependent tests

• queen-actors-dist: "a timer runs during a flood of 200 000 turns" times out at 5 s under load and under coverage.
• queen-actors: "a process takes its messages in order" fails under load. The first slice's closing setImmediate races the virtual clock's next timer.
• queen-actors-dist: "CPU turns in threads run side by side and leave the loop free" fails under load.
• pglive queen-actors-netlink-live: "a newer start on the same node id fences the old one" races the beat thread.
• The simulation gate: one seed's two runs yielded 637 and 634 times under coverage and load.

Done when

• The cost is bisected per merge and every free fix is made, with no change in behaviour. Each fix is proved with the bench: median of 21 rounds, run while the load average is below 6.
• Each test above is load-independent without weakening what it proves. It is reproduced under load first, then shown passing N times in a row under the same load.