Perf: publish quiescent owned batches without the allocation lock

Consumer parent: zig-utils/zig-js#97 Related: #25, #26, zig-utils/zig-js#125

Evidence

The post-#26 shared-heap profile shows metadata-lock handoff still dominates fixed-shape allocation. Two controlled alternatives have now failed:

• #25 shortened each ~17-cell critical section with an O(1) splice but regressed 2/4-lane throughput.
• zig-js#125 retained 64 or 256 initialized objects across checkpoints, but longer publication bursts regressed 4/8 lanes by 13–40%.

Changing exclusive-section length or frequency therefore cannot solve the convoy. Owned-slab heaps already classify managed pointers without the global list or payload map; while no mark is active, allocation only needs to make new headers discoverable for the next collection.

Proposed slice

• For parallel bindings that guarantee all cells use owned storage, initialize a non-marking batch privately and publish it onto an atomic pending-header stack without taking alloc_lock.
• Use an active-publisher counter plus the existing marking transition as a gate: a collector arms marking under alloc_lock, waits for pre-arm publishers to leave, folds the complete pending stack into all, then whitens/traces. A publisher that observes the transition falls back to the existing locked born-grey path.
• Aggregate live/byte/young accounting atomically once per fast batch. Collection threshold reads remain atomic in parallel mode; collection and teardown consume exact quiescent values.
• Retain the proven locked path for marking, concurrent born-cell handoff, non-owned storage, OOM prefixes, and single-mutator heaps.

Acceptance

• Prove no cell can publish between the collector's final pending drain and whiten/barrier arming.
• Preserve newest-first nursery prefix order, exact accounting, short-prefix/OOM ordering, and concurrent born-grey semantics.
• Add multi-mutator overlap tests that force publication/mark transitions and verify list/header/counter integrity under TSan.
• Pass the full zig-gc unit and TSan workflow.
• Exact-parent downstream A/B materially improves zig-js shared 4/8-lane object churn without regressing 1/2 lanes, direct, or independent contexts.

Rejected after complete implementation and downstream A/B; the dependency worktree is restored exactly and no commit was made. The active-publisher gate, pending atomic chain, aggregate counters, forced allocation/mark transition test, all 39 normal tests, and the full TSan test gate passed. Every zig-js checksum matched.\n\nSeven-sample ReleaseFast shared object_churn medians (100 jobs):\n- 1 lane: lock-free 132.207 ms vs parent 136.527 ms (-3.2%)\n- 2 lanes: 235.037 vs 237.682 ms (-1.1%)\n- 4 lanes: 579.181 vs 495.106 ms (+17.0%)\n- 8 lanes: 1,159.067 vs 1,029.700 ms (+12.6%)\n\nThe atomic pending-head and aggregate-counter cache lines become a new shared hotspot at 4/8 lanes. Together with #25 and zig-js#125, this rules out shortening the lock, lengthening batches, and replacing the lock with one global CAS lane. The remaining #97 work needs genuinely sharded/per-mutator metadata or less metadata per allocation, with collection consuming shards only at an existing rendezvous.