Calling `undefined` as a function silently no-ops instead of throwing TypeError

[proggeramlug]

Repro

const obj: any = {};
try {
  obj.missingFn(1, 2);
  console.log("did NOT throw");
} catch (e) {
  console.log("threw:", (e as Error).message);
}

Actual (Perry)

did NOT throw

Expected (Node)

threw: obj.missingFn is not a function

Impact

This is the single biggest leverage fix for the parity test suite. Calling a non-existent property as a function should throw TypeError: <expr> is not a function. Perry silently no-ops, which causes a cascading set of failures across the runtime parity tests:

1. test_parity_timers HANGS forever — the test does await new Promise(r => timers.setTimeout(cb, 0, 2, 3)). Since Namespace import for unrecognized node:* sub-paths resolves to true (TAG_TRUE) #629's partial fix, timers is an empty object instead of a boolean, so timers.setTimeout is undefined. The executor calls undefined(cb, 0, 2, 3) which Perry no-ops; resolve never fires; the await pends forever. Yesterday's binary threw TypeError: (boolean).setTimeout is not a function and the surrounding try/catch handled it cleanly. Today the test must be killed by the runner timeout.

2. Silent mid-test termination in many parity modules — test_parity_os finishes after 15 of 60 expected lines; test_parity_tls after 36 of 76; test_parity_perf_hooks after 43 of 84; test_parity_http2 after 62 of 137. In each case Perry runs into an undefined call mid-script and silently exits the surrounding async path without surfacing the error to the diff.

3. Confusing parity output — many user-visible bugs end up looking like "Perry produces fewer lines than Node" rather than the more honest "Perry threw at line X of test Y". A correct TypeError would be caught by per-section try/catch in the parity tests and surface as a one-line ERROR: row, dramatically improving the gap report quality.

Why this changed

Yesterday Perry NaN-boxed unrecognized node:* namespaces as TAG_TRUE (#629). Calling (true).foo() is a property access on a primitive boolean, which Node correctly throws on — and Perry's existing primitive-method-dispatch path threw too. After #629's partial fix flipped the namespace to an empty object, the path now goes through Perry's object-property lookup, which returns undefined. Calling undefined() is a separate code path that silently no-ops.

So the regression isn't in #629's change itself — it's that #629's fix exposed an existing latent bug in Perry's "call undefined as function" handling.

Acceptance

The 6-line repro above prints threw: obj.missingFn is not a function matching Node, AND test_parity_timers runs to completion without timing out (regardless of whether timers.setTimeout is actually implemented).

Related

• Indirectly enabled by Namespace import for unrecognized node:* sub-paths resolves to true (TAG_TRUE) #629's partial fix (namespace switched from TAG_TRUE → empty object).
• Cascades into ~10 of 38 parity tests that silently truncate mid-script.
• Same family as typeof process returns "number" instead of "object" (default-import + global) #623 (typeof returns wrong value for unrecognized binding) but on the call path, not the typeof path.

[proggeramlug]

Fixed in 3ef40d3 (v0.5.770). js_closure_callN (and _array / _apply_with_spread) and js_native_call_method's catch-all now throw <value>/<method> is not a function via js_throw_type_error_not_a_function, routed through Perry's exception system so try/catch catches it. The 6-line repro now prints threw: missingFn is not a function. Fallout: added explicit fs.rmSync HIR lowering since the catch-all previously masked it as a silent no-op. Sanity: 27/28 gap tests byte-identical to Node.