correctness: a declared-`number` + chain (number[] elements, `this` fields) reads all operands before the adds — the SECOND entry into the fused fold, not #10904's dynamic one

[proggeramlug]

A + chain over operands that are numeric only because an annotation says so is fused into one guarded tree, which reads every operand before any addition converts anything. A valueOf running in an earlier conversion can then change what a later operand reads, and perry adds the stale value.

This is the same hazard as #10904 reached through a different entry. #10904 is the dynamic add path (Any operands); this one is the declared-number path (numeric_proof_is_declared_only), and it is the spelling most real TypeScript has, so a reader auditing "is my code affected" will not recognise themselves in #10904.

Repro

// 1. element reads through a `number[]`
const A: number[] = [1, 2, 3];
(A as any)[0] = { valueOf() { A[2] = 100; return 1; } };
function sum(a: number[]): number { return a[0] + a[1] + a[2]; }
console.log("elements:", sum(A));        // node 103, perry 6

// 2. field reads through `this`, the commonest shape in class code
class H {
  a: number = 1; b: number = 2; c: number = 3; e: number = 4;
  sum(): number { return this.a + this.b + this.c + this.e; }
}
const H1 = new H();
(H1 as any).a = { valueOf() { H1.c = 100; return 1; } };
console.log("this fields:", H1.sum());   // node 107, perry 10

$ node --experimental-strip-types a.ts
elements: 103
this fields: 107

$ perry build a.ts -o a && ./a      # main 841b605c9 (v0.5.1632, train 252), Linux x86-64
elements: 6
this fields: 10

No crash and no diagnostic — just a wrong number.

(as any) is only how the repro installs the object; nothing about the bug needs it. Any value that reaches a number-annotated field or element while actually holding an object with a side-effecting valueOf/toString does it, which is exactly what annotations do not prevent (CLAUDE.md, Known Limitations: annotations are erased).

Why the operand order matters

For Add(L, R) the specification evaluates L, evaluates R, and only then ToPrimitives both. So in (a[0] + a[1]) + a[2] — which is what source-level a[0] + a[1] + a[2] parses to — a[0]'s and a[1]'s conversions run before a[2] is evaluated, and a valueOf in one of them can assign a[2]. The fold reads all three first.

The cold arm does not rescue it: rebuild_add_tree(.., fast = false) rebuilds over the already-loaded operand values, so even the spec-+ path adds the stale one. That is why the symptom is a wrong number rather than a crash.

Where it enters

lower_guarded_numeric_add (crates/perry-codegen/src/expr/binary.rs) is reached from two call sites:

1. the fully dynamic tree — gated by dynamic_add_tree_benefits_shared_guard, which fix(#10904): a + tree must not read a leaf after a conversion that precedes it #10921 taught to refuse an unfaithful tree;
2. both_numeric && numeric_proof_is_declared_only(..) — no check at all. A read whose "it is a Number" answer rests only on a declared type takes the guarded diamond precisely because the annotation can be wrong, and then the whole tree is fused unconditionally.

Entry 2 is this issue.

Scope

• Needs three or more operands with an inner left-leaning +. Two-operand chains are faithful by construction (both operands are evaluated before either is converted), and h + (a + b) is faithful too.
• Needs a read the compiler could not prove numeric by anything stronger than the annotation. A proven tier (Ptr<Shape> numeric field, element-shape loop fact, scalar replacement, typed array storage, POD record) answers false to numeric_proof_is_declared_only and never reaches the fold.
• Reproduced on main 841b605c9 (v0.5.1632, train 252), Linux x86-64. Not reproduced on the released 0.5.1220 on macOS arm64, which prints 103/107 — I have not bisected whether that is the version or the target.

Fix

Being fixed in #10921 by moving the faithfulness check to the top of lower_guarded_numeric_add, where every entry passes, rather than to either call site. A tree that would be read out of order lowers node by node through the spec helper. Fixture cases G (elements) and H (this fields) in test-files/test_parity_region_guards.ts pin both spellings, and a unit test pins the declared-number entry at the IR level.

This is worth knowing when reading matrix numbers: lane 13's read4__this__cctor/cfield cells got slower once this gate landed, because they had been taking the fused path and computing the wrong answer.

[proggeramlug]

Fixed on main. PR #10921 landed in merge train 254 as b3f71aee90 ("fix(#10904): a + tree must not read a leaf after a conversion that precedes it"), which states it fixes both #10904 and this declared-number entry.

Its close keyword never fired because trains rebase-merge, so the source PR is closed as landed rather than merged. Closing by hand.

Regression coverage for this specific spelling (number[] elements and number fields on this) is in #10959, riding merge train 256 (#11018). If that coverage shows the declared-number door is still open in some spelling, reopen rather than filing a new issue.