main no longer boots Claude-of-Duty: "Set.prototype.add is not a function" in world construction #9907

Summary

Current main no longer boots Claude-of-Duty. World construction dies with:

[boot] init failed TypeError: Set.prototype.add is not a function
TypeError: Set.prototype.add is not a function
    at solidSlabs        (src/world/util.js:591)
    at buildBuilding
    at WorldSystem.init
    at createConfig

Set.prototype.add is of course a builtin, so a receiver that reached this point is not being recognised as a Set.

The code

src/world/util.js:591 — nothing exotic:

export function solidSlabs(w, h, holes) {
  const xs = new Set([-w / 2, w / 2]);
  for (const o of holes) {
    xs.add(Math.max(-w / 2, o.x - o.w / 2));
    xs.add(Math.min(w / 2, o.x + o.w / 2));
  }
  const cuts = [...xs].sort((a, b) => a - b);
  // …
}

Regression window

That is 108 commits. Both bases have the root_reload fix, so it is not that.

What does NOT reproduce

Every reduction I tried passes on 890514a6d. Compiled and run individually:

1. new Set() + .add()
2. class Tracked extends Set overriding add and calling super.add
3. a Set held in a class field, mutated through a method
4. .add reached by computed key — (d as any)["add"](5)
5. new Set([1, 2, 3]) then .add(4) — literal elements
6. new Set([-w / 2, w / 2]) then .add(...) — computed elements
7. the whole solidSlabs body reproduced standalone, with holes
8. the game's real solidSlabs, imported from src/world/util.js and called directly — 75 MB binary, correct output

Number 8 is the informative one: the same function, same module, same import graph, compiled and called on its own, works. It only fails inside the full program.

That points at something context-dependent rather than at the lowering of this function in isolation — specialization driven by the real call sites (src/world/kit.js:106 passes openings) is my first guess, but I have not confirmed it.

Reproducer

git clone <claude-of-duty>
perry compile src/native/main.ts -o dist/cod
./dist/cod

Fails during world init, ~60 s in, before any window content.

I could not narrow it further without full game builds (~35 min each), which makes a 108-commit bisect impractical on my side. Someone able to bisect this range in CI should land on it quickly, and I am happy to test any candidate commit.

Note

An unrelated observation from the same comparison, in case it is a useful signal: the compiled binary grew from 106 MB to 151 MB across this same 108-commit window, with identical auto-optimize flags and near-identical runtime/stdlib archive sizes (26.7 -> 27.0 MB, 35.0 -> 35.3 MB). So the growth is in generated code.