perf: replace Activator.CreateInstance with direct constructor in ApiResponse.Create

[james-s-tayler]

Summary

Replace the reflection-based Activator.CreateInstance(typeof(ApiResponse<TBody>), ...) with a direct new ApiResponse<TBody>(...) constructor call in ApiResponse.Create<T, TBody>(), and change the content parameter type from object? to TBody? to eliminate boxing for value types.

This is a small, focused change that removes per-call reflection overhead on every ApiResponse<T> creation.

Why ApiResponse?

GitHub code search shows ApiResponse<T> is the second most common return type in Refit interfaces (after plain Task<T>):

• Task<ApiResponse<T>> found in 76+ unique repos (search cap hit at 100 file matches)
• Task<IApiResponse<T>> found in 47+ unique repos
• This is not a niche pattern — it's the recommended way to access HTTP status codes, headers, and error info alongside typed content

The typical Refit user profile (from GitHub search):

• Uses AddRefitClient<T>() with DI (97 repos)
• Uses System.Text.Json (default serializer, 91+ repos explicitly)
• Returns a mix of Task<T> and Task<ApiResponse<T>>

What Changed

Refit/ApiResponse.cs — 2 changes in the Create method:

1. Eliminate reflection: Activator.CreateInstance(typeof(ApiResponse<TBody>), ...) → new ApiResponse<TBody>(...)
2. Eliminate boxing: Parameter type object? content → TBody? content

Refit/RequestBuilderImplementation.cs — 1 change at the call site:

3. Fix null literal for unconstrained generic: null → default (required because TBody? for unconstrained generics doesn't accept null literal for value types)

Benchmark Results

BenchmarkDotNet v0.14.0, .NET 9.0.14, AMD Ryzen 7 7700 (8C/16T)

Baseline (before)

Scenario	Mean (μs)	Allocated
OK / Get / SystemTextJson	6.137	9.84 KB
OK / Get / NewtonsoftJson	9.105	19.52 KB
OK / Post / SystemTextJson	6.732	10.89 KB
OK / Post / NewtonsoftJson	14.204	32.03 KB
Error / Get / SystemTextJson	2.634	12.66 KB
Error / Get / NewtonsoftJson	2.549	12.66 KB
Error / Post / SystemTextJson	3.273	13.70 KB
Error / Post / NewtonsoftJson	7.561	25.16 KB

After optimization

Scenario	Mean (μs)	Allocated
OK / Get / SystemTextJson	5.949	9.52 KB
OK / Get / NewtonsoftJson	8.986	19.20 KB
OK / Post / SystemTextJson	6.669	10.56 KB
OK / Post / NewtonsoftJson	13.814	31.72 KB
Error / Get / SystemTextJson	2.350	12.33 KB
Error / Get / NewtonsoftJson	2.358	12.33 KB
Error / Post / SystemTextJson	3.111	13.38 KB
Error / Post / NewtonsoftJson	7.362	24.84 KB

Per-scenario improvement

Scenario	Δ Latency	Δ Allocations
OK / Get / STJ	-3.1%	-3.3%
OK / Get / Newtonsoft	-1.3%	-1.6%
OK / Post / STJ	-0.9%	-3.0%
OK / Post / Newtonsoft	-2.7%	-1.0%
Error / Get / STJ	-10.8%	-2.6%
Error / Get / Newtonsoft	-7.5%	-2.6%
Error / Post / STJ	-4.9%	-2.3%
Error / Post / Newtonsoft	-2.6%	-1.3%
Overall average	-3.1%	-1.9%

The error path sees the largest gains (-10.8%) since the reflection overhead is a larger fraction of total execution time on shorter code paths.

Consistent ~328 bytes/call allocation reduction across all scenarios from eliminating the boxing of TBody to object.

Commits

1. bench: add net9.0 target to Refit.Benchmarks — Adds net9.0 alongside net8.0 as a benchmark target
2. perf: replace Activator.CreateInstance with direct constructor — The optimization itself + call site null → default fix

Test results

All 501 tests in Refit.Tests and API approval tests pass on net9.0 with this change.