Hi! Thank you for looking into the issue and taking time to prepare a PR.
I was able to come up with another approach in #226. Instead of transforming e-notation numbers into custom JSON objects, it's more consistent to transform them into string interpolations, e.g. "${-5.1e2}".
Besides not introducing a custom structure and metadata, the practical advantage of that is terraform compatibility - let's consider floats.json file consisting of e-notation numbers saved in both formats:

{
 "floats": {
  "inline_scientific": [
   "${-5.1e2}",
   "${-3.7e1}"
  ],
  "object_scientific": [
   {
    "__sci_float__": true,
    "value": -510.0,
    "format": "-5.1e2"
   },
   {
    "__sci_float__": true,
    "value": -37.0,
    "format": "-3.7e1"
   }
  ]
 }
}

and try to do the below in terraform:

output "o1" {
  value = sum(jsondecode(templatefile("./floats.json", {}))["floats"]["inline_scientific"])
}

output "o2" {
  value = sum(jsondecode(templatefile("./floats.json", {}))["floats"]["object_scientific"])
}

templatefile call will read the file and automatically evalute string interpolations, in our case this results in "${-5.1e2}" string being evaluated into actual number. This means that the o1 output will be able to calculate sum of inline_scientific list, while o2 will fail, as object_scientific is a list of objects.
I understand the purpose of preserve_format flag, however this seems like a compromise we can avoid by using terraform-friendly string interpolations.