Cross-module instanceof returns true for unrelated imported classes #585

Repro

// lib_a.ts
export class A { ax = 1 }
export class B { bx = 2 }

// main.ts
import { A, B } from './lib_a.js';
const a = new A();
console.log(a instanceof A);  // true (correct)
console.log(a instanceof B);  // false (correct in this minimal case)

That minimal case actually works. But the bug surfaces when the value is an Object.assign'd instance of a class imported from one module, checked against an unrelated class imported from a different module:

// surfaces while compiling drizzle-orm 0.45.2 via perry.compilePackages
function pgTableWithSchema(name) {
  const rawTable = new PgTable(name);          // PgTable lives in pg-core/table.js
  return Object.assign(rawTable, { /* cols */ });
}
const users = pgTableWithSchema('users');

// drizzle-orm/entity.js's `is()` brand check
import { Subquery } from '../../subquery.js';   // unrelated class, separate module
console.log(users instanceof Subquery);          // perry: TRUE (WRONG); node/bun: FALSE

In the actual drizzle run via perry, all three of these return true:

• users instanceof Subquery
• users instanceof PgViewBase
• users instanceof SQL

Why it matters for #488

drizzle-orm's is(value, kind) brand-check (in entity.js) short-circuits with if (value instanceof type) return true;. With this bug, is(table, Subquery) returns true unconditionally. drizzle's from() builder then takes the Subquery branch, reads table._.selectedFields, and crashes with TypeError: Cannot read properties of undefined (reading 'selectedFields').

This is the dominant query-side blocker for #488 (drizzle + @perryts/postgres acceptance test).

Where I'd start digging

crates/perry-codegen/src/expr.rs:5498 Expr::InstanceOf — for the cross-module case, falls back to cid = 0 when ctx.class_ids.get(ty) misses (line 5526-5539). Then js_instanceof(value, 0) is called.

crates/perry-runtime/src/object.rs:3806 js_instanceof — at line 4047-4050:

let obj_class_id = (*obj_ptr).class_id;
if obj_class_id == class_id { return true_val; }

If both obj_class_id and the requested class_id happen to be 0 (e.g. an Object.assign'd table whose class_id metadata was lost + an unresolved cross-module class name), 0 == 0 returns true.

Probable fix surfaces:

1. Codegen: ensure cross-module Subquery/PgViewBase/SQL resolve to a non-zero class id (likely working — see imported_class_stubs registration in codegen.rs:462). Verify they're being threaded through correctly when compilePackages is used vs. ordinary .ts imports.
2. Runtime: never return true for class_id == 0 — that should be an "unknown class" sentinel that always returns false.
3. Codegen: Object.assign(rawTable, { ... }) may lose the rawTable's class_id; verify the result still has the correct ObjectHeader.class_id.

Workaround

Patch drizzle's entity.js to skip the instanceof short-circuit and rely on the entityKind chain walk:

// if (value instanceof type) return true;  // skip on perry

With this patch the chain walk also fails (A[sym] returns undefined for static Symbol-keyed fields — separate issue), but the path advances to the next blocker (#TBD: thenables). So is() is the gate.

Refs

Blocks #488. Surfaced from #420 sub-issue acceptance run on v0.5.684.