perf: optimize search processing

[devlsh]

Summary

- Repeated frontier scans and string cell IDs
+ Indexed frontier and sparse numeric cell records

• Move internal helpers into domain modules.
• Add BENCH_BASELINE for comparisons against another checkout.
• Breaking change: remove package exports calculatePath, neighbors, score, vectorId, and asc. Keep search and consumer types.

Evidence

Historical search-latency change against the original main baseline, before later cleanup:

Group	Change
Quick	-86.26%
Full	-93.14%

Median across five processes per group. Host contention limits precision. These are previous results, not a fresh measurement of this PR head.

Passed on the pushed head: pnpm check, pnpm test (29 tests), pnpm build, and the local native CLI protocol (22 assertions).

AI-assisted implementation. Independent Standards and Spec reviews informed the changes.

Merge Danger

Door: two-way

Blast Radius: consumers

Existing imports of the removed helper exports require migration. Review search compatibility and the export changes.