test(benchmarks): cover the CurrentSpeedBenchmark performance gate

[DemchaAV]

Summary

Adds the missing test for CurrentSpeedBenchmark#evaluatePerformanceGate — the only benchmark check that actually fails a run. It enforces the smoke-profile absolute thresholds (avg latency + peak heap per scenario); previously it had zero coverage.

What's covered

CurrentSpeedBenchmarkPerfGateTest drives the gate with synthetic LatencyRow values, so the verdict is deterministic and machine-independent (the gate is not run against real measurements, which would be flaky):

• passes when every scenario is within thresholds
• fails on average-latency breach (message names the scenario)
• fails on peak-heap breach
• aggregates failures across multiple scenarios (and leaves passing ones out)
• skips the gate for non-smoke profiles
• ignores scenarios that have no configured threshold
• honors the per-scenario graphcompose.benchmark.gate.<scenario>.maxAvgMillis override

Production change

Minimal test seam only — no behaviour change:

• evaluatePerformanceGate -> static package-private (it uses no instance state)
• LatencyRow, PerformanceGateResult, BenchmarkProfile -> package-private

so the same-package test can build inputs and read the verdict (same pattern as the existing BenchmarkDiffTool.ResolvedRunPair seam).

Verification

./mvnw -f benchmarks/pom.xml test -> 11/11 green (7 new perf-gate + 2 selection + 2 median); the new class runs in ~30 ms.

Note

Independent of #133 (different file: CurrentSpeedBenchmark vs BenchmarkDiffTool); both target develop and do not overlap.