`(() :> string)` compiles, and other coercions and variant spreads treat `bool` and `unit` as variants

Start with Variant_coercion.can_try_coerce_variant_to_primitive and type_is_variant in compiler/ml/variant_coercion.ml, then trace subtype_rec in Ctype and constructor spreading in compiler/ml/variant_type_spread.ml. Reproduce the coercion and spread cases from the issue, then run the compiler tests that cover variant coercions and spreads. Done means bool and unit values retain correct runtime behavior and invalid coercions are rejected; pull request #9218 is already open.

unit can be coerced to string, a closed polymorphic variant whose tags are "true", "false" or "()" can be coerced to bool or unit, and bool or unit can be spread into a variant. The coercions compile to nothing, so undefined gets the type string, and a string gets the type bool or unit. The compiler also suggests these coercions: let s: string = u() for u: unit => unit is reported with "These types are compatible at runtime. You can use the coercion operator to convert to the expected type: (u() :> string)", and let b: bool = p for p: [#"true" | #"false"] with (p :> bool).

let s: string = (() :> string)
let n = s->String.length

let n = undefined.length;

let s;

Loading the module throws TypeError: Cannot read properties of undefined (reading 'length'). The subtype check accepts unit for string wherever it compares the two: (f :> string => string) for f: string => unit, (g :> unit => int) for g: string => int, (p :> promise<string>) for p: promise<unit>, and a unit in a tuple, a record or object field or a constructor payload.

let toBool = (p: [#"true" | #"false"]): bool => (p :> bool)
let answer = toBool(#"false") ? "yes" : "no"
let isFalse = toBool(#"false") == false

function toBool(p) {
  return p;
}

let isFalse = "false" === false;

let answer = "yes";

toBool(#"false") returns the string "false", which is truthy and is not false, so answer is "yes" and isFalse is false. In the same way, let u: unit = (p :> unit) for p: [#"()"] makes u the string "()", and u == () is false.

A variant that spreads bool builds true and false, but its switches and coercions take them for the strings "true" and "false":

type t = | ...bool | Other
let x: t = true
let s = (x :> string)
let r = switch x { | true => "yes" | false => "no" | Other => "other" }

let r;

switch (true) {
  case "false" :
    r = "no";
    break;
  case "true" :
    r = "yes";
    break;
  case "Other" :
    r = "other";
    break;
}

let x = true;

let s = true;

r is undefined, and s is the boolean true typed as a string. On master, a variant that spreads unit builds undefined for (), and its switch compares undefined === "()"; 13.0.0-alpha.6 and earlier build the string "()" there.

Some coercions of unit that are rejected give reasons that do not apply to it. (() :> int) says "The constructor () of variant unit has no @as payload, which makes it a string at runtime", ("a" :> unit) says "Fix this by adding the @unboxed attribute to the variant unit", and on master (() :> unknown) says that the variants have different @unboxed configurations.

bool and unit are declared in compiler/ml/predef.ml as variants with the constructors false, true and () and no @as, but their values are false, true and undefined. Variant_coercion.can_try_coerce_variant_to_primitive in compiler/ml/variant_coercion.ml leaves out bool but not unit, so Ctype.subtype_rec checks unit with variant_has_same_runtime_representation_as_target, which takes a constructor without @as to be its name as a string, and so accepts string. Variant_coercion.type_is_variant, which guards the coercion from a polymorphic variant to a variant, leaves out neither type, and can_coerce_polyvariant_to_variant then matches #"true" to the constructor true by name. Variant_type_spread.map_constructors in compiler/ml/variant_type_spread.ml copies the constructors of bool and unit like those of any variant; Translcore.transl_exp builds true and false by name as booleans, and on master Lambda.const_constructor builds () as undefined, while the spread variant matches and coerces them as untagged constructors, which are strings.

Reproduced with 12.0.0-alpha.6, 12.3.1, 13.0.0-alpha.6 and master (c10223e80). 11.1.4 and 12.0.0-alpha.1 also accept (() :> string) and the spreads, but reject the polymorphic variant coercions, which became possible with #6981. The suggested coercion and the @as and @unboxed reasons appear in 12.3.1, 13.0.0-alpha.6 and master, not in 12.0.0-alpha.6.

Related: #6311, #6316, #6980, #6981, #7505.