Perf: shard cooperative GC allocation-pressure accounting

Parent: #97 Follows: zig-utils/zig-gc#33

Evidence

Both exact owned-enumeration candidates retained one zig-js-wide write on every managed allocation batch: GcCellBacking.parallel_cell_bytes_since_collection.fetchAdd. The counter only decides when shared workers have issued the next cooperative-GC tranche, while allocation already holds the matching size-class bucket lock.

The first traversal-only activation improved four lanes 30% but not eight. The second candidate also removed zig-gc's heap allocation lock/global aggregate writes per large batch; it improved 1/2/4-lane medians 4.8% / 1.5% / 8.1% but regressed eight lanes 6.0%.

Evaluated design

• replace the global allocation-byte RMW with per-size-class counters updated under the bucket lock already held by allocation;
• sum/reset every counter under all bucket locks at the cooperative safepoint;
• preserve exact reset/threshold semantics;
• leave the tranche, wait primitive, sweep/unpublication, batch size, and collector binding unchanged.

Result

The isolated exact-parent A/B regressed every median: 0.95x / 0.87x / 0.87x / 0.84x at 1/2/4/8 lanes. It lost every paired sample at four and eight lanes. The candidate and its tests were therefore reverted. Raw results and the decision are preserved in docs/.data/object-churn-pressure-accounting-ab-2026-07-18.md.

Acceptance

• Normal and suppression-free TSan tests proved concurrent increments, exact sums, and reset exclusion for the evaluated candidate.
• Cooperative shared-nursery lifecycle tests remained exact and leak-free.
• The evaluated allocation hot path performed no heap-wide atomic RMW for tranche accounting.
• Exact-parent shared object_churn A/B failed the 4/8 gate, so the candidate was reverted as required.
• README benchmark scores remained unchanged.

Decision

The global atomic remains. Any successor needs a cheap non-blocking read path and exact reset semantics before repeating the same gate.

Implemented, race-tested, measured, and reverted under the issue's acceptance rule.

Exact isolated A/B (parent/candidate median speedup):

• 1 lane: 0.95x (3/7 paired wins)
• 2 lanes: 0.87x (3/7)
• 4 lanes: 0.87x (0/7)
• 8 lanes: 0.84x (0/7)

Normal and suppression-free TSan backing/cooperative tests passed before the measurement. Because the candidate regressed every median and lost every 4/8 pair, retaining it would be a net regression.

Evidence landed in 96b643dd. README scores are unchanged.