Integrate inference smoke diagnostics

internal/generation/evidence.go

@@ -18,6 +18,7 @@ type GenerationEvidence struct {
 // AttemptEvidence describes one capacity acquisition and any dispatch it
 // reached. Generated content remains on the separate response return value.
 type AttemptEvidence struct {
+	StartedAt                 time.Time
 	Dispatched                bool
 	ResponseReceived          bool
 	HTTP                      llm.HTTPExecutionEvidence