fix(hir): a later class accessor replaces an earlier one - #9886

The bug

ECMA-262 ClassDefinitionEvaluation installs class elements in source order, so a second get x / set x replaces the first. Perry keeps the first.

ClassDecl::getters / ::setters are consumed with iter().find(...) — first match wins — but every accessor was appended with push(). The shadowed definition stays live and the one the program actually defines last is silently dropped.

Reproducer

class Vec {
  a = 111;
  c = 222;
  damped = 0;

  set z(v: number) { this.damped = v; }
  get z(): number { return this.a; }   // shadowed
  get z(): number { return this.c; }   // must win
}

const v = new Vec();
v.z = 7;
console.log(v.z, v.damped);

Note the setter still runs in every case: when a later definition supplies only a getter, the earlier setter stays in force. That asymmetry is the point — see below.

Every accessor shape is affected

Measured against Node on the fixture in the regression test:

There is no diagnostic. The program reads a plausible value from the wrong accessor and keeps running.

How it surfaced

Claude-of-Duty's Spring3 (src/weapons/mathx.js) pairs an early set z (damping) with a later get z (displacement):

export class Spring3 {
  constructor(f = 12, z = 1) { this.a = new Spring(f, z); /* … */ }
  set z(v) { this.a.z = this.b.z = this.c.z = v; }  // damping in
  get z()  { return this.a.z; }                     // shadowed
  // …
  get z()  { return this.c.x; }                     // displacement out — must win
}

Unusual, but legal, and it depends on exactly the read/write asymmetry the spec produces. Perry served the shadowed damping getter, so:

lag.z    = 0.4600   ← new Spring3(5.4, 0.46) damping
recPos.z = 0.4200   ← new Spring3(9,   0.42) damping

Those are added as displacements, contributing +0.880 m to the first-person viewmodel's Z. The rig moved from its authored hipPos.z = -0.300 (in front of the eye) to +0.580 (behind it):

rig.pos = 0.117, -0.186, +0.580
basePos = 0.118, -0.185, -0.300

All 156 viewmodel nodes then sat at camera-space Z ≥ 0 and clipped. The failure was invisible from the renderer's side: the overlay pass ran, the gate passed, all 156 draw_indexed calls were issued, and the pass's colour reached the screen — with zero fragments. lag.x and lag.y were correct throughout, because only .z collides with the damping accessor.

The change

record_class_accessor overwrites an existing entry instead of appending.

The replacement is keyed on (name, is_static), not on the name alone: a static and an instance accessor may legally share a name and are distinct properties — one on the constructor, one on the prototype. Deduping on the key alone would trade this bug for another, which is why the regression test includes a class carrying both.

Verification

• cargo test --release -p perry-hir — 620 passed, 0 failed
• cargo test --release -p perry-codegen — 1912 passed, 0 failed
• New test duplicate_class_accessor_last_wins covers all five shapes above, with expected output taken from Node. Reverting the one-line lowering change fails it on every shape:

left:  "spring 111 7\nsplit instance-first static-first\nsink first:x 1\nexpr 1\n"
right: "spring 222 7\nsplit instance-last static-last\nsink last:x 1\nexpr 2\n"

Summary by CodeRabbit

• Bug Fixes

  • Fixed duplicate class accessors so later getter or setter definitions correctly replace earlier ones.
  • Preserved separate behavior for static and instance accessors with the same name.
  • Applied consistent accessor handling to class declarations and class expressions.

• Tests

  • Added regression coverage verifying duplicate accessor resolution and runtime behavior.