One folder, one bounded context, in each language
Yes, one bounded context can be one folder with its own project, in all four languages. Each was built and run with 2 to 3 contexts, a kernel and tests inside the folder, then scaled to 25 or more. You meant project, and that is what was built: a standard project, not one library per layer.
The folder, as files
Explorer
- context.yamlid, owner, needs, uses
- what others may use
- registerthe one function the kernel calls
- URLs, controllers, request and response models
- MCP tools
- model, service, ports
- data model, repository, migrations
- policies.yamlauth policies
- its own settings
- flags.yamlrequired
- segments.yamloptional
- events.yamloptional
- contract, alerts, runbooks
- its own docs
- finds and wires every context
- dependency.yamlone per service
What each language allows
| Language | What a context is | Files outside the folder when you add one | What blocks another context's internals |
|---|---|---|---|
| Go | one package tree, one Go module | 1 generated wiring file | the compiler |
| Bun | one workspace package and a contract package | 1: bun.lock | TypeScript; our checker for relative paths |
| .NET | a project, a contract project, a test project | 0, with no lock file in the host | the compiler |
| Flutter | three packages: contract, src, tests | 0 tracked | only our checker |
All MEASURED.
Can tests, docs and observability live inside?
| Language | Tests inside run | Planted failing test fails | Test files in the shipped build |
|---|---|---|---|
| Go | yes | yes | 0 |
| Bun | yes, with SIT | yes | 0 |
| .NET | yes | yes | 0 |
| Flutter | yes | yes | 0 |
- Docs and runbook files were kept out of every shipped build. MEASURED
- The obs tool ran on a context's own folder in Bun and Flutter. MEASURED. Go and .NET: NOT TESTED.
- The docs framework was not run on these folders: it is not finished. NOT TESTED
- SIT ran only in Bun. Flutter device tests were NOT TESTED.
Guarding the layers inside one project
The compiler no longer does it. A check does, and each was proven with a planted breach. MEASURED
| Language | Check | When it fails |
|---|---|---|
| Go | depguard and our generator | lint, generate |
| Bun | our 150-line boundary checker | check |
| .NET | our analyzer | build |
| Flutter | our boundary checker | check |
Speed with 25 contexts
| Language | From clean | After one edit | One context's tests |
|---|---|---|---|
| Go | 43 s | 2.2 to 4.3 s | 0.75 s |
| Bun | 5.5 s (type check) | 3.7 s | 0.21 s |
| .NET, the app | 77 s | 10 s | 2.8 s |
| .NET, with every test project | 161 s | 11.7 s | 2.8 s |
| Flutter, 27 contexts | 43 s (generate, analyze) | 8.3 s | 6.8 s |
MEASURED on a busy 32-core box. Compare within a language, not between languages.
Does this change .NET the most? Yes. Is it possible? Yes.
| Rank | Language | Template files that move | What really changes |
|---|---|---|---|
| 1, most | .NET | 64 of 428 (API template) | the four layer projects and the solution file go; the shared DbContext; a new layer check |
| 2 | Flutter | 0 of 559 (template has no features); 180 in alcohol.neon | packages per context; two generated files; a boundary check |
| 3 | Bun | 42 of 409 (consumer template) | packages per context; 6 root configs; a boundary check |
| 4, least | Go | 4 of 248 | about 10 config files; one generated file per context |
- Counts: MEASURED, except Flutter's 0, which is REASONED. The order is REASONED from structure, because the templates differ in size.
- One page each: Go, Bun, .NET, Flutter.
What each language cannot do cleanly
- Go: no
src/folder and no module per context. Two new libraries on neighbouring lines conflict ingo.mod. - Bun: two new contexts conflict in
bun.lock, one hunk, fixed bybun install.merge=unionis unsafe there. - .NET: a lock file in the host conflicts on every new context. Keep none there, as the template does today.
- Flutter: one shared package map, so an undeclared import compiles. Stale generated files break every command until
task gen. Two new libraries can conflict inpubspec.lock. - All of these lock-file clashes are git conflicts, so none breaks main silently.