Regression: ZenStack v3 does not properly generate TypeScript string literal discriminators for delegated models

[ellie-xpawn]

Description and expected behavior
After migrating from ZenStack v2 to v3, there is no longer a type-safe way to discriminate TypeScript types generated from delegated models.

model Base {
  id String @id @default(cuid(2))
  type String
  @@delegate(type)
}

model Foo extends Base {
  foo Number
}

model Bar extends Base {
  bar String
}

The types generated are as follows:

type Base = {
  id: string
  type: string // Expected: "Foo"
  foo: number
} | {
  id: string
  type: string // Expected: "Bar"
  bar: string
}

type Foo = {
  id: string
  type: string // Expected: "Foo"
  foo: number
}

type Bar = {
  id: string
  type: string // Expected: "Bar"
  bar: string
}

It's the same with enums as well - if an enum is used in place of String, then every submodel still will not narrow types as expected:

enum Type {
  Foo
  Bar
}

model Base {
  ...
  @@delegate(Type)
}
...

type Base = {
  id: string
  type: "Foo" | "Bar"
  foo: number
} | {
  id: string
  type: "Foo" | "Bar"
  bar: string
}

type Foo = {
  id: string
  type: "Foo" | "Bar"
  foo: number
}

type Bar = {
  id: string
  type: "Foo" | "Bar"
  bar: string
}

Because of this, Base cannot be narrowed to Foo or Bar without opting out of type checking and using a type assertion.

This didn't work entirely as expected in v2 either, but workarounds were possible. See zenstackhq/zenstack#2334 for more info.

Screenshots
If applicable, add screenshots to help explain your problem.

Environment (please complete the following information):

• ZenStack version: 3.2.1
• Database type: PostgreSQL

Additional context
Add any other context about the problem here.