feat: per-queue resource awareness with measured CPU/memory estimates

[sylvesterdamgaard]

Summary

Resolves #14 — Phase 1 + Phase 2 foundation for per-queue resource-aware capacity calculations.

• New ResourceEstimate value object — carries per-worker CPU cores and memory MB estimates with per-dimension source metadata (Measured, Config, Default) and sample counts
• New ResourceEstimateResolver singleton — three-source precedence chain: measured runtime data > per-queue config override > global default. Each dimension (CPU, memory) resolves independently
• CapacityCalculator refactored — calculateMaxWorkers() now requires a ResourceEstimate parameter instead of reading from internal state; memory calculations are now adaptive (were previously static config only)
• AutoscaleManager writes per-queue measured estimates — replaces the old updateMeasuredWorkerCpuEstimate() (global weighted average across all jobs) with updateMeasuredResourceEstimates() that groups by connection:queue and tracks both CPU and memory
• Per-queue resources config — operators can declare cpu_cores and memory_mb per queue as cold-start fallbacks before measured data is available

What this enables

For heterogeneous workloads (e.g. fast 50MB jobs vs slow 2GB jobs on different queues), the autoscaler now uses the actual resource shape of each queue for capacity math instead of flattening everything to a single global average. This prevents over-provisioning fast queues and under-provisioning heavy queues.

Architecture

ResourceEstimateResolver (singleton)
  ├─ AutoscaleManager → writes per-queue measured data each tick
  ├─ ScalingEngine    → reads resolved estimate before capacity call
  └─ CapacityCalculator → pure function of (workers, estimate, system metrics)

Resolution order per dimension: measured → queue.resources config → limits.worker_*_estimate

Not included (future phases)

• Phase 3: Resource-aware bin-packing in distributeClusterTarget() — the resolver and per-queue estimates are in place, distribution just needs to consume them
• Confidence-based fallback (e.g. ignore measured if sampleCount < 50)

Test plan

• 525 tests passing (1473 assertions)
• PHPStan clean (1 pre-existing property.unusedType warning, not introduced by this PR)
• Pint clean
• New unit tests for ResourceEstimate, ResourceEstimateResolver (11 tests covering all precedence paths, clamping, partial measurement, queue isolation, reset)
• Integration tests verifying heterogeneous workloads produce different capacity results
• All existing CapacityCalculator, ScalingEngine, policy, cluster, and simulation tests updated and passing