fix: prevent worker pool thrashing in cluster distribution layer

[sylvesterdamgaard]

Summary

Fixes worker pool thrashing where current_workers oscillated ±2 every few seconds despite a stable cluster-wide target. The autoscaler was killing and respawning workers in rapid succession due to sort-order instability in the distribution layer.

• Root cause: distributeClusterTarget() sorted managers by live heartbeat worker counts ($currentCounts), which fluctuate tick-to-tick as workers spawn/die. Different sort orders produced different per-host assignments each cycle, even when the cluster-wide target was unchanged.
• Fix: Added distribution hysteresis — cache previous per-workload host assignments and reuse them when the target sum, active manager set, and per-host capacity are all unchanged. Fresh computation only triggers when the target actually changes, a manager joins/leaves, or capacity constraints shift.
• Tests: 6 new tests including a 30-cycle stability invariant that replays fluctuating heartbeat counts and asserts identical assignments across all cycles.

Closes #21

Test plan

• 30-cycle stability invariant: stable target with fluctuating heartbeat counts produces identical assignments every cycle
• Recomputes distribution when target changes (scale up/down)
• Recomputes when a manager joins the cluster
• Recomputes when a manager leaves the cluster
• Recomputes when cached assignments exceed host capacity (other workloads consumed slots)
• All 550 existing tests pass — no regressions
• PHPStan: no new errors
• Pint: formatted

🤖 Generated with Claude Code