Untagged variants: no warning for patterns shadowed by a literal case, and a spurious non-exhaustive warning for bool payloads #8806

Summary

An untagged variant may have a literal case whose runtime value a payload case can also hold (#8632, #8777):

@unboxed type t = | @as("a") A | S(string)

Matching tests declared literals first, so the pattern S("a") can never match. The value "a" always goes to A, or to a wildcard if there is no A arm. The type checker treats S("a") and A as different, which causes two problems.

1. No warning for patterns that can never match

Checked on master. None of these produce a warning, and each S(...) arm marked dead is dead in the generated JS:

let f1 = x => switch x { | A => 1 | S("a") => 2 | S(_) => 3 }  // S("a") dead
let f2 = x => switch x { | S("a") => 2 | A => 1 | S(_) => 3 }  // dead even as first arm
let f3 = x => switch x { | S("a") => 2 | _ => 3 }              // compiles to `return 3`
let f4 = x => switch x { | S("a" | "b") => 2 | _ => 3 }        // the "a" alternative is dead

The same applies when:

• the literal comes from the constructor name (Primary | Color(string): Color("Primary") is dead);
• an int literal overlaps a float payload (@as(1) One | F(float): F(1.0) is dead);
• a nested untagged constant has the same value (@as("x") X | W(inner) where inner has @as("x") P: W(P) is dead).

In a refutable binding such as let l = (S("a") as v) => v, the pattern never matches at all, so every call with S("a") throws Match_failure.

Proposal: warn on a constructor pattern S(p) of an untagged variant when every value p can match is the runtime value of a literal case in the same variant. Use warning 11 (unused case) for a whole arm and warning 12 (redundant sub-pattern) for or-pattern alternatives. Suggested message: "this pattern can never match: "a" is the value of constructor A".

Notes:

• The check is local to each pattern and doesn't depend on arm order, so it can run beside check_unused without touching the exhaustiveness algorithm.
• Values must be compared the way they are at runtime, as in Fix constant folding of overlapping unboxed variant matches #8631: 1 and 1.0 are equal, bigint spellings differ (1_0n = 10n), ints use their 32-bit value, and a constructor without @as is represented by its own name.
• Guards (S(s) if s == "a") are out of scope.

2. Spurious non-exhaustive warning for bool payloads

With #8777 a bool payload can partly overlap a literal:

@unboxed type b = | @as(true) Yes | Ans(bool)
let g = x => switch x { | Yes => 1 | Ans(false) => 2 }

This match covers every value, yet the compiler reports warning 8 with the counterexample Ans(true) and emits an unreachable Match_failure throw. The only way to silence it is to add the dead arm Ans(true) => ..., which point 1 would then flag. The two parts therefore need to land together, or the bool case ends up with no warning-free way to write the match.

Proposal: during the exhaustiveness check only, treat a literal-case pattern as also covering the payload values equal to it, e.g. check Yes as Yes | Ans(true). This is harder than point 1 because it changes parmatch.ml. For string and number payloads it rarely matters, since their domain is infinite, but for bool it does.

Related: #8632, #8777, #8631, rescript-lang/rescript-lang.org#1337