GeezerVison

Summary

Improve startup, search, memory use, and HTTP navigation for large libraries:

• Tokenize repeated words and lines once, avoid redundant fragment parsing and retained copies of search text, and update indexes incrementally after content edits.
• Prepare search indexes and vault checks in batches so other actions can proceed; skip unnecessary legacy Notes asset scans at startup.
• Virtualize HTTP navigation and eliminate repeated tree allocations and per-node document scans.
• Add isolated benchmark vaults, measurement tools, and a shared HTTP development server; remove the superseded demo servers.
• Allow the perf conventional commit type.

Measurements

Internal synthetic-library tests on an Apple M4 Pro with approximately 30,000 records:

Startup results use 36 alternating original/current build launches across two library sizes and three spaces, with identical profile/vault copies and a warmed filesystem cache. Search and filtering compare historical sessions; memory uses a focused before/after experiment. Search measures API processing, filtering measures list updates, and startup ends at the selected item's open event. These are not input-to-screen measurements. Results vary with content and hardware.

Validation

• 218 scoped tests passed during review, plus two regression tests for HTTP benchmark readiness. Both regression tests fail on the previous implementation and pass after the fix.
• Scoped ESLint, main TypeScript and production renderer builds passed; diff checks are clean.
• Completed 36 startup measurements and six HTTP runtime remeasurements after correcting the readiness gate.
• Reviewed storage/search/API and renderer/Electron/benchmark changes. The additional HTTP benchmark readiness finding is fixed.

Separate follow-up

The known drag-source lifecycle and scroll-preservation fixes remain in fix/tree-drag-scroll and are intentionally excluded from this PR. Benchmark reports and release drafts are local, gitignored artifacts.