react-reconciler: flushSyncWork → commitRoot calls forward-declared hoisted `flushPendingEffects` as undefined (next wall after #5015)

[proggeramlug]

Summary

After #5019 (queueMicrotask/structuredClone/atob/btoa as first-class function values) fixed #5015, react-reconciler's updateContainerSync succeeds, but the next call — flushSyncWork() — throws TypeError: value is not a function. This is a distinct bug from #5015: it reproduces even with an arrow scheduleMicrotask (so it is independent of the queueMicrotask value fix).

This is the next wall on the ink end-to-end path (#348), immediately after #5015.

Root cause (localized)

The non-callable is a 0-arg call inside commitRoot in react-reconciler.production.js:

function commitRoot(root, finishedWork, …) {
  root.cancelPendingCommit = null;
  do flushPendingEffects();          // <-- flushPendingEffects reads back as `undefined`
  while (0 !== pendingEffectsStatus);
  …
  flushMutationEffects();
  flushLayoutEffects();
  flushSpawnedWork();
}

flushPendingEffects (and its siblings) are forward-referenced, hoisted sibling function declarations, defined after commitRoot inside the giant module.exports = function($$$config){ … } factory, and captured by commitRoot as boxed (mutable) captures. At the time commitRoot runs (during flushSyncWork), the captured box for flushPendingEffects still holds undefined — i.e. the function-declaration's box assignment and the captured box have desynced (or the late assignment never reached the captured box).

Confirmed via --debug-symbols + lldb: the failing js_closure_call0 site is the do flushPendingEffects() call (the instruction reads js_closure_get_capture_f64(this, N) → js_box_get → 0x1/undefined), immediately after the cancelPendingCommit = null js_put_value_set.

Repro

import React from 'react';
import createReconciler from 'react-reconciler';
const noop = () => {};
const rec: any = createReconciler({
  supportsMutation: true, supportsMicrotasks: true,
  getRootHostContext: () => ({}), getChildHostContext: () => ({}), getPublicInstance: (i: any)=>i,
  prepareForCommit: () => null, resetAfterCommit: noop, clearContainer: noop,
  createInstance: () => ({ children: [] }), createTextInstance: () => ({ text: '' }),
  appendInitialChild: noop, finalizeInitialChildren: () => false, shouldSetTextContent: () => false,
  appendChild: noop, appendChildToContainer: noop, insertBefore: noop, removeChild: noop, removeChildFromContainer: noop,
  commitUpdate: noop, commitTextUpdate: noop, detachDeletedInstance: noop, getCurrentEventPriority: () => 0,
  maySuspendCommit: () => false,
  // independent of #5015: an arrow here still reproduces
  scheduleMicrotask: (cb: any) => queueMicrotask(cb), scheduleTimeout: setTimeout, cancelTimeout: clearTimeout, noTimeout: -1,
  preparePortalMount: noop, prepareScopeUpdate: noop, getInstanceFromScope: () => null, getInstanceFromNode: () => null,
  beforeActiveInstanceBlur: noop, afterActiveInstanceBlur: noop,
});
const container = rec.createContainer({ children: [] }, 0, null, false, null, 'id', ()=>{}, ()=>{}, ()=>{}, null);
function App() { return React.createElement('host', null, 'hi'); }
rec.updateContainerSync(React.createElement(App), container, null, noop);  // OK after #5019
rec.flushSyncWork();   // → TypeError: value is not a function (flushPendingEffects undefined)

(compilePackages: ["react","react-reconciler","scheduler"], perry.define NODE_ENV=production.)

Notes

• Minimal hand-written repros of "earlier sibling fn-decl forward-references a later one, captured + escaped, called in a do-while, with 25+ captures" all currently pass — the trigger appears tied to the scale/structure of the compiled-package factory (thousands of statements / hundreds of nested function declarations) or the compiled-package lowering path. A faithful reduction is still needed.
• Same family as compilePackages: cross-module mutable object loses identity — react-reconciler's ReactSharedInternals.H dispatcher invisible to react → 'Invalid hook call' null dispatcher (blocks all React renderers) #4950 / Captured Object.assign alias causes react-reconciler HostRoot render hang #2564 (cross-module / captured callable identity).
• Relevant code: crates/perry-hir/src/lower_decl/body_stmt/nested_fn_decl.rs (nested fn-decl → Stmt::Let { Closure }, capture/mutable-capture detection + HIR: function-scoped function declarations not hoisted — silent runtime undefined #569 hoisting) and the codegen box-allocation/sharing for hoisted function-decl locals.

Related

• react-reconciler 0.33 production build: updateContainerSync throws 'value is not a function' on first render (minimal 30-line repro, no ink) — blocks all React renderers #5015 / fix(hir): queueMicrotask/structuredClone/atob/btoa as first-class function values (#5015) #5019 (the wall this one is immediately behind), Compile ink (React-based TUI framework) end-to-end via perry.compilePackages #348 (ink end-to-end)

[proggeramlug]

Cannot reproduce on current main — closing as resolved

Reproduced and traced this end-to-end (worktree off fresh origin/main, native compilePackages build of react-reconciler driving updateContainerSync → flushSyncWork).

1. The repro in the issue body is itself invalid. Its host-config object is missing methods react-reconciler reads from $$$config (notably getCurrentUpdatePriority/resolveUpdatePriority). Node throws on it too:

TypeError: getCurrentUpdatePriority is not a function
    at commitRoot (.../react-reconciler.development.js)

The "value is not a function" symptom from the minimal repro is this config gap, not a codegen bug. Under --debug-symbols + lldb the failing 0-arg js_closure_call0 site is getCurrentUpdatePriority(), not do flushPendingEffects().

2. With a complete (node-valid) config, the flushPendingEffects path works on Perry. On react-reconciler@0.32.0 (the version that introduced the exact flushPendingEffects function this issue names), a full config that runs cleanly under node --experimental-strip-types also compiles + runs correctly under Perry:

OK: flushSyncWork completed

flushSyncWork → commitRoot → do flushPendingEffects() executes to completion; output matches node. Same result on 0.31.0.

3. The boxing machinery this issue blamed is consistent on HEAD. Instrumenting the compiler confirms every Stmt::PreallocateBoxes id is present in the module-wide module_boxed_vars (prealloc_missing_from_modbox=0), so there is no "captured box vs. declaration box" desync: the hoisted sibling function is prealloc'd, boxed, and its box is js_box_set with the closure before any capturing closure runs. The earlier-captured box and the declaration's box are the same cell.

Net: the reported flushPendingEffects-box-desync does not occur on current main — resolved by the #5019-era work plus everything since (#5042/#5165/#5171, …). Closing as not-reproducible. Happy to reopen with a self-contained repro that fails under Perry but passes under node --experimental-strip-types (i.e. a node-valid config).