PR #16027: actual base/head Linux web comparison
Base: 687bd55533, original Pierre patch hash298e94372dbd51846a4e26ece2ecc3e22eb1f94b0450b1f1a9c6fec127809be7.
Head candidate patch hash2ec81ba1e8ce4681095a3473c2bf78e630417740c622720ef624febb467c25fc.
Actual main dependency patch and lock were installed for baseline. Final files restored byte-for-byte and pnpm install repeated for head. Same isolated app state and fixture,1280x800 viewport,all files collapsed,scroll to last header.
Fixture:96 real Git files,400 replaced lines each,+38400/-38400. No provider turns.
Instrumented actual browser Worker.prototype.postMessage, counting only message.type === "diff"; returned original call unchanged. Instrumentation installed before opening DiffPanel after reload; zero existing diffs-container elements.
Base: after passive scroll/load to file-096.ts,96 diff-highlight worker requests,0 expanded files.
Head: after passive scroll/load to file-096.ts,0 diff-highlight worker requests,0 expanded files.
Head find-in-diff: SCROLL_PROOF_NEEDLE_096 =>1 of1,real rendered matching code,only file-096.ts unfolded;1 diff-highlight request.
No frame-time/FPS benchmark. Video is2x playback; intended cost shifts to opening a file and syntax highlighting can arrive after its plain text.
Focused Chromium suite:36/36 in3existingfiles (main34, original38); baseline collapsed AST/highlight counts1/2 become0/0. Scope lint/fmt/webtypecheck passed. Independent ClaudeOpus5.5 and CodexGPT6Sol reviews both clean in round1.
No macOS,Windows,Electron desktop,iOS,Android or remote/relay/tunnel integrated verification.