fix: bound hierarchical scan scratch ranges

[SamJSui]

Summary

• bind hierarchical scan data and auxiliary storage views to exact logical byte ranges
• add a 4,194,305-item regression that forces a third high-end hierarchy level
• publish the previously withheld 100M scan and compaction comparisons on RTX and both Jetsons
• update the roadmap now that the correctness blocker is closed

Root cause

Deeper hierarchy levels place data and auxiliary at different offsets in one scratch buffer. Both bindings previously used size: None, so the data view extended through the next auxiliary level. Rounded-up WGSL threads used that larger arrayLength(&data) and could overwrite the next level, producing a scheduling-sensitive writable-storage race.

The fix bounds data to num_items * 4 bytes and auxiliary to ceil(num_items / items_per_block) * 4 bytes. It changes views only; it does not add allocation or copies.

Validation

• all three captured 100M scan and compaction processes pass on RTX, 8-TPC dopey, and 4-TPC grumpy
• 10M controls changed by -0.25% to -1.39%; no measured regression
• cargo fmt --all --check
• cargo test --lib --tests
• cargo clippy --all-targets --all-features -- -D warnings
• cargo doc --no-deps
• cargo package --allow-dirty
• JSON parse and git diff --check

Performance finding

Correctness is closed, but valid 100M data confirms the next optimization target. Massively is 1.95x faster for scan and 1.49x for compaction on RTX, and 2.03x/1.71x faster on the 4-TPC Jetson.