Related to #24531, but a distinct trigger — see the comparison below.

Repro

export function f(a) {
	if (a) var x = true;
	else var x = false;
	return x ? "ok" : "fail";
}

minify("t.js", code, { compress: true, mangle: false }) →

export function f(a){if(a)var x=!0;else var x=!1;return`fail`}

f(true) returns "fail" instead of "ok" — the true from the first
branch is discarded. Notably the swapped order compiles correctly:

if (a) var x = false;
else var x = true;        // → x ? "ok" : "fail" is preserved ✓

which suggests the analysis models the binding as
{undefined, <last declaration's initializer>}, dropping the earlier
declaration's initializer: {undefined, false} is uniformly falsy → folded;
{undefined, true} is mixed → not folded.

How this differs from #24531

• No closure involved. minifier: var declared with a initializer conditionally is inlined unconditionally #24531 needs the var to be captured by a
  returned arrow function; here the condition is folded in plain
  straight-line code within the same function. The single-declaration shape
  from minifier: var declared with a initializer conditionally is inlined unconditionally #24531 compiles correctly without the closure — this repro fails
  even without one.
• The var is redeclared in both branches of an if/else; minifier: var declared with a initializer conditionally is inlined unconditionally #24531 has a
  single declaration in one branch.
• The dropped value is an actual assignment (the first branch's
  initializer), not the implicit undefined of a never-taken branch as in
  minifier: var declared with a initializer conditionally is inlined unconditionally #24531. Consequently the fold goes the other way: minifier: var declared with a initializer conditionally is inlined unconditionally #24531 folds to the
  assigned value, this one folds to falsy even though one branch assigns
  true.

Both look like the same underlying reaching-definitions defect (one reaching
value ignored at the use site), so a fix for #24531 may or may not cover
this — worth verifying against both repros.

Real-world impact

Scala.js routinely emits this pattern for branch-dependent locals. In our
production app it hit the CBOR validator of the
borer library, whose onBreak method
compiles to exactly this shape — compress deleted the valid-BREAK path and
made it unconditionally throw, so every CBOR message containing an
indefinite-length array failed to decode in vite production builds while
dev builds (unminified) worked.

Versions

• oxc-minify 0.140.0 (standalone, latest at time of writing)
• rolldown 1.1.4 and 1.1.5
• vite 8.1.3 (default production minify); vite 7.x (esbuild minify) unaffected

Workaround

minify: { compress: false, mangle: true, removeWhitespace: true }