Document smoke-test diagnostics

docs/internal/capacity.md

@@ -88,6 +88,17 @@ release, including canceled acquisition without dispatch. A retry wait holds
 the logical admission lease but no active permit. Observer callbacks must not
 block on work that requires the same run's admission slot.
 
+## Smoke-Test Capacity
+
+The engine's smoke tester shares this manager and generation coordinator with
+the ordinary runner. Inference diagnostics reserve the same logical admission
+slot and acquire the same backend-scoped active permits for actual model
+attempts, including retries. They therefore participate in the same configured
+backend limits as `Run`. Reachability and authentication HTTP probes do not
+enter generation admission or consume active permits. The public
+[`Engine.SmokeTest` GoDoc](../../smoke_diagnostics.go) owns the caller-visible
+operation and injected-client limits.
+
 ## Cancellation And Release
 
 Admission checks the caller context before reserving a slot. A call canceled
@@ -127,4 +138,7 @@ cover shared and independent backend pools during retry and repair. The
 own the public prepared-capacity boundary. The
 [root error-boundary tests](../../errors_internal_test.go) own preservation of
 the public generation category and context identity when generation is
-canceled.
+canceled. The assembled
+[smoke diagnostic contract tests](../../smoke_integration_contract_test.go)
+verify that inference diagnostics share the engine's Run admission and recover
+capacity after cancellation or failure.