This PR fixes a regression introduced in #10917.

In https://github.com/babel/babel/pull/10917/files#diff-dd23409d66e27b24936f7ae4e316073eR265, the optional meta-field is added only if the default meta-field presents. This change makes sense but unexpectedly it breaks

which relies on optional to detect if it is a nullable key, so the non-builder fields are marked as non-optional and included in the definition. For example, here is what Program definition

generates in different versions.

v7.7.4

export function program(
  body: Array<Statement>,
  directives?: Array<Directive>,
  sourceType?: "script" | "module",
  interpreter?: InterpreterDirective | null,
  sourceFile?: string | null
): Program;

v7.8.0

export function program(
  body: Array<Statement>,
  directives: Array<Directive> | undefined,
  sourceType: "script" | "module" | undefined,
  interpreter: InterpreterDirective | null | undefined,
  sourceFile: string
): Program;

The sourceFile is not nullable so all the parameters before sourceFile becomes non optional parameters. This is a breaking change for types consumers.

The problem here is that we should not include sourceFile at all because it is not defined in builders and we are really lucky that nobody ever complains that the fifth argument sourceFile? does not work at all.

This PR filter the args against BUILDER_KEYS so the generated types become

export function program(
  body: Array<Statement>,
  directives?: Array<Directive>,
  sourceType?: "script" | "module",
  interpreter?: InterpreterDirective | null
): Program;

where the sourceFile is removed from builder arguments definition, technically it is a breaking change but practically it should not break anyone because the builder will check the arguments length at runtime.

In the long term we should add the generated index.js.flow and index.d.ts to the codebase so we can track such changes.