#846 · zig-utils/zig-js

Parent: #461 Discovered by: #844, #845

Problem

The native representative attribution gate deterministically rejected shared 2-lane representative_weak_collections at warmup. Backing accounting was exact (60,614,174 added - 36,184,922 released = 24,429,252 current), and allocation-kind counts decomposed exactly, but GC-cell lifetime bytes did not: 28,463,893 acquired < 33,871,815 released.

Two source paths were omitted:

GcCellBacking.allocateCellBatch returned fresh or reused owned slots without calling the runtime attribution profiler. Later sweep paths did record their frees.
Delegated GC-side storage resize/remap growth and shrink changed the eventual free size without updating acquired/released lifetime byte totals.

Scope

Attribute every successfully returned batched owned slot with its exact bucket size and fresh/reused origin.
Attribute successful delegated resize/remap growth and shrink deltas without manufacturing allocation/free operation counts.
Preserve the existing batch lock boundary, parallel allocation-pressure accounting, and allocator semantics.
Do not estimate from requested aggregate sizes or weaken the gc_cell_bytes >= gc_cell_freed_bytes publication invariant.
Add focused mixed fresh/reused, partial-batch, resize, and remap lifetime tests.

Acceptance

Batched fresh and reused allocations increment total count/bytes and exactly one kind per returned slot.
Partial batches account only slots actually returned.
Successful delegated growth/shrink/remap updates byte lifetime while allocation/free counts remain exact.
Batched and delegated releases never exceed attributed lifetime acquisition bytes.
Focused Debug and TSan tests pass.
Representative attribution quick collection completes end to end with all existing allocation gates unchanged.

Security boundary

Profiler evidence is fail-closed. The repair accounts the real allocator lifecycle at its source; suppressing the invariant or clamping released bytes would conceal allocator/GC defects.