test(benchmarks): align Polly comparisons

[thomhurst]

Summary

• align Kevlar and Polly circuit-breaker benchmarks on the same ratio, sampling, throughput, and break settings
• compare manually isolated circuits on both sides and token-bucket rate limiters on both sides
• make retry delays explicitly zero wherever only strategy machinery is intended to be measured
• alternate and aggregate stress-test rounds so runner drift does not systematically penalize the second library
• give corrected scenarios new history keys so previous non-equivalent measurements are not mixed into published medians

Motivation

The 15-minute stress run measured Kevlar about 5.6% below Polly despite the published microbenchmarks favoring Kevlar. The stress workload itself used matching ratio breakers, while the published composed benchmarks compared Kevlar's consecutive-failure breaker with Polly's ratio breaker. The rate-limit comparison also used Kevlar's token-bucket/GCRA limiter against Polly's fixed-window limiter, and the stress harness always measured Polly first and Kevlar second.

This change makes the compared semantics explicit and symmetric and balances early/late runner conditions.

Local benchmark check

BenchmarkDotNet short job, .NET 10.0.11, Windows, Intel Core Ultra 9 185H:

Scenario	Kevlar	Polly	Allocation
Ratio breaker, closed	270 ns	315 ns	0 B / 24 B
Timeout → retry → ratio breaker	454 ns	1,236 ns	0 B / 48 B
Five-strategy token-bucket/ratio chain	572 ns	1,203 ns	0 B / 88 B
Token bucket, uncontended	152 ns	188 ns	0 B / 0 B

The short job is directional; the normal benchmark workflow remains the authoritative measurement.

Validation

• dotnet build Kevlar.slnx -c Release
• dotnet run --project tests/Kevlar.Tests -c Release --no-build -- --timeout 5m — 685 passed
• dotnet run --project tests/Kevlar.AllocationTests -c Release --no-build -- --timeout 5m — 3 passed
• BenchmarkDotNet dry validation — all 88 cases passed
• BenchmarkDotNet corrected 13-case dry and short runs passed
• 8-second alternating-round stress smoke test and both documentation generators passed