perf: fuse compaction block prefixes

[SamJSui]

Summary

• keep compaction offsets block-local instead of materializing global offsets
• add the scanned preceding-block total during stable scatter
• remove one full-size offsets read/write pass for both u32 and KeyValue compaction
• add explicit portable no-subgroup coverage and publish fresh RTX/Jetson evidence

Why

At 100M items, compaction spent about 1.35 ms traversing the 400 MB offsets buffer solely to add block prefixes. Scatter immediately reread those values. The fused path removes that pass (800 MB of logical read/write traffic) while preserving standalone scan output.

Massively 0.96 results

Speedup is Massively time divided by wgpu-primitives time.

System	Compact 10M	Compact 100M
RTX 4070 Ti SUPER	2.06x	1.52x
dopey, 8-TPC Orin	1.66x	1.52x
grumpy, 4-TPC Orin	1.63x	1.49x

Versus PR #17, compaction improves 19.9%-27.3%. Fresh exact-current 10M matrices also retain Jetson leads of 8.48x-8.55x for bounded sort, 4.53x-4.55x for full-width sort, and 1.48x-1.53x for scan. The complete fresh RTX matrix wins all 12 sort/scan/compaction rows by 1.12x-19.71x.

Full method and process medians: benchmarks/2026-08-08-fused-compaction-prefix.md.

Validation

• cargo fmt --all --check
• cargo test --lib --tests (64 passed)
• cargo clippy --all-targets --all-features -- -D warnings
• cargo doc --no-deps
• cargo package --allow-dirty
• portable compaction with SUBGROUP disabled at 4,097 items
• 26 scoped release scan/compaction/profiling tests on each Jetson
• every process-isolated 10M/100M benchmark passed correctness readback
• Jetson scan controls within 0.21%; sort controls within 0.38%
• both Jetsons restored and verified on dynamic CPU/GPU/EMC ranges with nvfancontrol active