perf(core): lazily allocate row geometry index

[chh-ay]

Summary

• represent uniform row geometry arithmetically instead of eagerly allocating and building two million-entry typed arrays
• materialize the dense height/Fenwick backing stores only on the first custom row height or non-default inserted band
• retain exact offset, structural edit, custom-height, scaled-scroll, and allocation-error behavior
• record backingStoreBytes in the million-row diagnostic contract and refresh the three-round evidence

Measurement

• direct 1M-row constructor baseline: 5.871 ms median, 16,000,008 typed-array bytes
• optimized direct constructor: 0.000270 ms median, 0 backing-store bytes
• retained browser diagnostic: prior 6.90-7.11 ms/operation; now 0.000119-0.000125 ms/operation across three rounds
• all 12 diagnostic cells passed; formula, long-text, and fractional-scroll scenarios remained within their prior absolute ranges

Verification

• bun test packages/core/test: 765 passed
• focused core/bench tests: 29 passed
• repository bun run typecheck: passed
• Biome check on changed source: passed
• retained diagnostic replay: 12/12 passed
• changeset validation against 0.4.0: passed

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
sheetwrite	Ready	Preview, Comment	Jul 27, 2026 3:44pm