docs: the vue-nodes perf record's numbers cannot be reproduced from the committed scripts

## Summary

`docs/architecture/vue-nodes-large-graph-performance.md` presents a Result table and a headline DOM-element count that no committed script can produce, so a reader cannot check the numbers and a future maintainer cannot re-run them.

## What was checked

Round 4 ran a citation-verification pass over the document. Two corrections to expectations first, in the document's favour: it contains **no** PR or issue numbers and **no** external URLs, so there are no miscitations. The gap is evidentiary, not citational.

## The gaps

1. **No committed script produces the Result table.** `scripts/perf/gen-sized-workflow.mjs`, `measure-culling.mjs`, `measure-zoom.mjs` and `sweep-threshold.mjs` are all committed and runnable-looking, but none of them emits the combination of graph sizes, zoom levels and renderer set that the table reports.
2. **The headline DOM counts are in a unit nothing measures.** "227,787 → 5,355 → 418 DOM elements" is a count of all elements. The committed scripts count `[data-node-id]` node roots, never `'*'`, and they run a hardcoded 245-node fixture with no zoom-out step, so neither the 3000-node figure nor the zoomed-out figure has a source.
3. **The Method block mischaracterises `sweep-threshold.mjs`.** Tracked separately: that script sets `ViewportCulling` to `true` under a comment saying it forces culling off, so as committed it cannot produce both arms of the comparison it is described as producing.
4. Several quantitative claims in prose have no committed source, and the variance caveat is self-contradictory: variance cannot be characterised from one sample per configuration.

## Why it is worth fixing rather than deleting

The document is unusually thorough for a perf change and the reverted-approach history in it is genuinely valuable. The problem is narrow: the numbers and the scripts drifted apart. Either extend the scripts to emit exactly the table, or mark the table as a point-in-time manual measurement and state the environment (DPR, viewport, build mode) so a reader knows what they are comparing against.

## Affected area

- `docs/architecture/vue-nodes-large-graph-performance.md`
- `scripts/perf/*.mjs`

Found during round 4 review of #15031.