Perf: right-size the object-only GC slab class

[chrisbbreuer]

Parent: #62
Follow-up to rejected #78.

Evidence

Compile-time Heap.cellAllocationBytes values on current main:

• object: 656 bytes (624-byte payload + 32-byte header)
• function: 288 bytes
• generator: 264 bytes
• environment: 248 bytes
• promise: 192 bytes
• iterator helper / module namespace: 112 bytes
• bound function: 64 bytes

The current slab classes are 64, 128, 256, 512, 1024, 2048, so ordinary objects alone cross the 512-byte boundary and consume a 1024-byte slot: 368 bytes / 35.9% internal waste per object. #78 added a 768-byte seventh class and regressed object_churn about 9.8%, consistent with adding another linear classification/ownership bucket to every hot path.

Proposed slice

• replace the existing 1024-byte class with an exact 656-byte object class instead of adding a bucket;
• keep the bucket count and classification depth unchanged;
• retain the large-chunk policy for the 656-byte object class;
• derive the regression test from Heap.cellAllocationBytes(Object) so layout drift fails visibly;
• update allocator docs that name the large object classes;
• preserve fallback/delegation, ownership indexes, free-list reuse, trimming, bulk teardown, heap caps, and parallel locking.

Acceptance

• compile-time proof still shows every managed cell maps to owned slab storage;
• focused GC backing/nursery tests pass with zero leaks;
• GC profile confirms ordinary cells use the 656-byte class with fewer/equal chunks and correct live/free accounting;
• seven alternating current-main/candidate process pairs preserve the exact object_churn checksum and show a meaningful median win;
• only then run the batched full suite/harness and land a small conventional commit.

[chrisbbreuer]

Rejected after two controlled six-class A/B variants against b20d30e5; no code retained.

First, focused validation exposed a build-mode layout constraint: ReleaseFast needs 656 bytes per object cell, while Debug needs 672. The replacement therefore must be at least 672 bytes to preserve the full matrix.

672-byte replacement

• kept six buckets and the 384 KiB large-chunk policy;
• 18/18 focused backing tests passed with zero leaks;
• ReleaseFast profile improved from 3 chunks / 1152 slots to 2 chunks / 1170 slots;
• alternating single object_churn 100 3: 0/7 pair wins; median 232.274250 -> 246.750041 ms (+6.23% time); all 42 checksums exact.

768-byte replacement

• also kept six buckets, isolating it from Perf: add a dense GC slab class for ordinary objects #78's seventh-bucket search cost;
• alternating A/B: 0/7 pair wins; median-of-process-medians 231.618125 -> 259.356583 ms (+11.98% time); all 42 checksums exact.

The existing 1024-byte power-of-two stride is materially faster despite its internal fragmentation. Future #62 work should reduce allocation/sweep operations or shrink Object enough to enter the existing 512-byte class; tightening the current slab stride is not viable. The worktree and docs are fully reverted.