fix(codegen): PreallocateBoxes must not shadow a module-level global (#7521) - #7580

Closes #7521.

Root cause

Not in diagnostics_channel, and not a closure-capture bug. traceCallback is
a bystander — the minimal repro has no traceCallback in it, and no array:

import { tracingChannel } from "node:diagnostics_channel";   // any import at all
{
  const events: string[] = [];
  function t1(a: any) { events.push("fn:" + a); }
  t1("A");
  console.log(JSON.stringify(events));   // node: ["fn:A"]   perry: []
}

Objects, numbers and strings lose the same way — inside t1, the captured
binding simply reads undefined, so acc.n++ throws
Cannot read properties of undefined and events.push(...) silently
no-ops. The import matters only because it makes the file an ES module,
hence strict, which is what selects the affected lowering. Everything else
about the failing gap test (traceCallback, subscribers, arrows, the array)
was incidental.

crates/perry-codegen/src/stmt/mod.rs, emit_preallocate_boxes: it allocated
a heap box — and, decisively, a ctx.locals slot — for every id in a
Stmt::PreallocateBoxes directive, including ids that
codegen/module_globals_emit.rs had already promoted to
@perry_global_<mod>__<id>.

Codegen already has a rule for this collision, spelled out identically at every
read and write site in expr/literals_vars.rs and stmt/let_stmt.rs:

ctx.boxed_vars.contains(id) && !ctx.module_globals.contains_key(id)

The module global wins — it is already the shared, forward-visible, GC-rooted
cell a box would provide. emit_preallocate_boxes was the one place that never
checked, and the stale ctx.locals entry it left behind is consulted before
ctx.module_globals on both the Stmt::Let reuse path (let_stmt.rs:322) and
the LocalGet/LocalSet store paths (literals_vars.rs:642, :814). So the
declaration wrote its value into the local box-pointer slot, the module global
was never stored, and the closure — which reads through the global, with zero
captures — saw the undefined the global was defined with.

The emitted IR, before:

%r2  = call i64 @js_box_alloc_bits(i64 <undefined>)   ; PreallocateBoxes([1])
%r8  = call i64 @js_closure_alloc(ptr @perry_closure_m3_ts__0, i32 0)
%r14 = call i64 @js_array_alloc(i32 0)                ; const events = []
store ... %r16, ptr %r3                               ; -> the box-pointer slot

@perry_global_m3_ts__1 is defined, registered as a GC root, read and written
by the closure — and never stored by main.

Why now

c6ed8175d (#6853) is not the window. lower_strict_block_fn_decls was added
by #7105 (76fa7b4c9, 2026-08-01), which made a bare block at the top level
of any ES module emit PreallocateBoxes for its let/const bindings. That is
the first time a promoted module-level id was ever handed to
emit_preallocate_boxes. The two commits the issue records as PASS
(17c0ff952, c6ed8175d, both 2026-07-30) both predate it, so the recorded
evidence is consistent — no bisect was needed once the diff was in view.

Fix

Skip promoted ids in emit_preallocate_boxes. The global is statically
initialized to TAG_UNDEFINED, which is exactly what a non-TDZ prealloc box
seeds, so nothing is lost. The TDZ variant is skipped too and the reason is
written into the code: module-global reads are a raw load double @g with no
js_box_get_bits choke point, so seeding TAG_TDZ there would leak the
sentinel into arithmetic instead of throwing a ReferenceError — strictly worse
than the undefined a forward read gets today.

Tests

crates/perry-codegen/src/stmt/prealloc_module_global_tests.rs — an in-src
unit test module, deliberately not crates/*/tests/*.rs, so it runs in the
per-PR cargo-test gate forever rather than only nightly (CLAUDE.md #5960).
The gap test that surfaced this (test_gap_diagchannel_3082_3084_3085_3086.ts)
is in neither gap_snapshot.json nor known_failures.json and parity is
tag-gated, which is how this sat unnoticed since 2026-07-30.

Four tests, and they are sabotage-checked rather than merely green — with the
ctx.module_globals guard reverted:

The store assertion counts store … ptr @perry_global_… lines inside main()
specifically. A plain contains was not enough and was caught doing nothing:
main also takes the global's address for js_gc_register_global_root, which
is emitted whether or not the declaration ever writes the cell, so the first
draft of that test passed under sabotage.

Validation (local — CI has a deep backlog and may not report)

• test_gap_diagchannel_3082_3084_3085_3086.ts byte-matches
  node --experimental-strip-types (Node 26.5.1), exit 0, all four blocks
  (node:diagnostics_channel: forward all runStores callback arguments #3082, node:diagnostics_channel: reject symbol tracingChannel names #3084, node:diagnostics_channel: preserve non-callable bindStore transform errors #3085, node:diagnostics_channel: honor traceCallback position and arguments #3086) — not just the one that was failing.
• 5 hand-built reduction probes (traceCallback, plain channel, no channel at
  all, function decl / function expression / arrow, array / object / number /
  string accumulators) all match node exactly.
• cargo test -p perry-codegen --lib — 671 passed, 0 failed.
• cargo test -p perry-hir --lib — 281 passed, 0 failed.
• RUST_TEST_THREADS=1 cargo test -p perry-runtime --lib — 1812 passed, 0 failed.
• Parity sweep over every gap test containing a module-top-level bare block
  (the exact shape this touches) plus a broader gap slice — all pass.
• cargo fmt --all -- --check clean; scripts/check_file_size.sh OK;
  scripts/raw_handle_debt.py 998 (baseline 998, unchanged).

Scope note

The blast radius is wider than one gap test: every ES module with a
top-level { … } block containing a function declaration that reads a
sibling let/const compiled to a silently-empty binding. Nothing threw, so
this class of breakage was invisible except where a test happened to print the
accumulated value.

https://claude.ai/code/session_019EHcmXKArA7m42SihYCcgH

Summary by CodeRabbit

• Bug Fixes

  • Fixed module-global bindings being incorrectly allocated as local boxes, preventing empty or invalid bindings and preserving correct global access.
  • Ensured temporal dead zone handling remains correct for promoted module globals.
  • Preserved heap-box allocation for ordinary function-local captured bindings.

• Tests

  • Added regression coverage for module-global and captured-binding behavior.
  • Removed an outdated known-failure entry.

• Chores

  • Updated the documented and package version to 0.5.1328.