Most of the work never needed to happen.

Opusghost used to bundle everything into one file and load all of it before the first page appeared. Word import and export, ZIP handling, analytics and recording were all in that file, whether you used them today or not.

Production builds are now minified local ES modules with named lazy boundaries, so tree shaking still works and the optional features load on demand. Nothing was moved to a CDN — a stick with no network gets exactly the same app.

Not a claim that the offline cache is 453 KB. The site precaches every chunk on first load, so Word and recording work offline without ever being opened while online. The desktop app keeps all of them on disk for the same reason.

To draw one page of the live manuscript as a turning sheet, Opusghost deep-cloned the entire document and then deleted every block that was not on that page. On a 5,000-paragraph manuscript that was 20,002 elements created and destroyed per page turn, to keep 34 of them.

The clone now skips the descendants it was only going to delete. The resulting snapshot is the same page. Sheet shells, which cloned the whole editor just to empty it, went from 4,005 elements to 3.

Clone timings in the actual system webview: medians over nine alternating before-and-after trials, 256 clones each. These are not end-to-end frame-rate claims. Animation keyframes, easing, strip counts, durations and typography are untouched.

To stay under its 2 MB cap, the recorder dropped one event at a time and re-serialised the whole buffer with JSON.stringify to check whether it still fitted. That is quadratic, on the renderer’s main thread, every 500 events — while you are typing.

It now finds the range to discard in a single pass and copies the survivors once. The live recorder sizes each event as it arrives and evicts incrementally, so the trim no longer runs on the keystroke path at all.

The first three buffer sizes in the original table were all 1 ms. The cost only appeared once the buffer crossed the cap, which is exactly why it survived so long. Output was verified identical by running the old implementation beside the new one over 21,000 randomised cases: zero mismatches, zero mutated inputs.

On these meters the before column is the recorded baseline from the work that fixed the bug, which cannot be re-measured without reverting the fix. The after column was re-measured on the current build, so small differences against the originally recorded figure are run-to-run variance. Both are batch-trim or append-plus-snapshot timings on synthetic events — not typing latency, and not process-RAM measurements.