[erikshafer]

What this is

An experiment, not a slice: slice 001 (POST /cakes, GET /cakes, GET /cakes/{id}) rebuilt on the Clean Architecture Solution Template (dotnet new ca-sln, version 10.8.0), run against the same contract scenarios as a third host, and counted. It exists to test the slide-2.5 claim that the before twin is "the way the templates generate" rather than reword it.

It lives in experiments/before-clean-template/ and tests/LayerCake.ContractTests.CleanTemplate/, neither of which is in LayerCake.slnx. dotnet build and dotnet test at the root are exactly what they were on main; nothing under src/, openspec/, docs/slices/, or the shared scenario classes changed. The scorecard sections of docs/file-inventory.md are untouched; the experiment has its own section at the end, and docs/build-log.md has the full narrative.

The number

Same three endpoints, .cs files created / edited:

Host	Created	Edited	Touched
Before twin	19	4	23
Template	9	8	17 (13 if the four Guid-key ripple edits count as a one-time cost)
After twin	5	1	6

Running it

dotnet test tests/LayerCake.ContractTests.CleanTemplate/LayerCake.ContractTests.CleanTemplate.csproj

Docker running is the only prerequisite (Testcontainers PostgreSQL, same fixture as the twins). Result as committed: 6 of 10 green (ping plus five cake scenarios). The four error scenarios (two 400s, 404, 409) return the right status codes and are red on content type alone: the template's exception handler writes application/json, never application/problem+json. The two 400 bodies also carry no field names; the template builds a ValidationProblemDetails and serialises it as the base type, so the errors dictionary is dropped. One line in the template's handler (Results.Problem(problemDetails).ExecuteAsync) makes it 10 of 10; that was probed and reverted so the branch shows the template as shipped.

The template host was also run live and driven from the repo's demo page; every route answered as the suite says.

Why the third host is a separate test project

Adding the template to LayerCake.ContractTests unified the test bin to MediatR 14.1.0 and AutoMapper 16.1.1, and every before-twin scenario failed with a TypeLoadException before running. One process, one MediatR. The separate project references the existing suite for the unchanged scenario base classes and aliases the template's Web assembly because all three hosts expose a public top-level Program.

Scaffold edits (outside the count)

• global.json floor lowered to the installed 10.0.1xx SDK band.
• NU1902;NU1903 appended to the template's own WarningsNotAsErrors: as generated on 2026-09-06 the template fails restore on NuGet audit warnings promoted to errors.

Open calls

• Keep the branch at the honest 6 of 10, or apply the one-line problem+json fix as a listed scaffold edit.
• Whether the four Guid-key ripple edits count as feature edits (17) or scaffold (13).
• Whether an Ardalis variant (before-clean-arch) is worth about an hour.