Perf: batch owned-cell unpublication during sweep

Downstream: zig-utils/zig-js#97

Evidence

The current sweep releases equal-size storage in 64-cell batches, but it first calls unpublishCellAllocation once per dead cell. In zig-js that acquires the same size-class lock for every bitmap removal, then freeCellStorageBatch reacquires it once for the matching free-bitmap run. The shared object_churn workload produces long runs of same-size dead Object cells, and the latest eight-lane sample still shows serialized sweep/backing-bitmap work after publication batching.

The atomic-header-guard experiment in #31 did not pass the downstream gate. This issue keeps exact bitmap ownership and instead batches the existing lock-protected state transition.

Design

• Add an optional unpublishCellAllocationBatch binding hook with a per-cell fallback.
• In sweep, identify a bounded consecutive run of same-size white cells, unpublish the complete run under one binding lock, then finalize, unlink, unindex, poison, and release it in the existing order.
• Never batch across an alive cell, allocation size, young-generation boundary, or collection boundary.
• Preserve the invariant that no cell is finalized or returned to reuse while its owned-storage classifier still reports it published.

Acceptance criteria

• Exact owned-storage tests prove a same-size dead run uses one unpublication batch and remains unclassified before reuse.
• Mixed live/dead types and minor/full collection accounting and finalization order remain exact.
• Full zig-gc normal suite and focused zig-js parallel-GC semantic gate pass.
• On an order-balanced exact-parent zig-js object_churn A/B, do not regress 1/2 lanes and materially improve 4 and 8 lanes; otherwise fully revert and document the result.

Rejected after the exact downstream performance gate; the prototype has been fully reverted in both repositories.

Correctness before measurement:

• full zig-gc suite passed;
• exact backing-lock proof passed 2/2;
• focused zig-js parallel-GC gate passed 5/5 with no leaks;
• every benchmark checksum matched the exact parent.

Order-balanced seven-pair ReleaseFast zig-js shared object_churn A/B, 100 jobs:

Although the implementation removes per-cell ownership-bitmap locking during sweep, it fails the explicit no-regression/material-improvement contract at both high-contention lanes. This reinforces that ownership bitmap transitions are not the governing 4/8-lane bottleneck; the next experiment needs to remove or parallelize collector traversal/global publication work rather than further batch the existing sweep tail.

No README score was changed because this was not an accepted full matrix.