The article contains no proof of this statement.

If someone knows of research that proves this, please let me know.

I think this satisfies the threshold of "minimum viable publication". So write this up and reference it here.

Full impl.:

getLineStart(tokenPosition, anteTerminators = null) {
  if (tokenPosition > this._edge && tokenPosition != this.length) {
    throw new Error("random access too far out"); // XXX
  }

  // NB: Since line terminators can be the multibyte CRLF sequence, care
  // must be taken to ensure we work for calls where `tokenPosition` is some
  // start minus 1, where that "start" is some line start itself.
  for (let i = this._lineTerminators.length - 1; i >= 0; --i) {
    let current = this._lineTerminators[i];
    if (tokenPosition >= current.position + current.content.length) {
      if (anteTerminators) {
        anteTerminators.push(...this._lineTerminators.slice(0, i));
      }
      return current.position + current.content.length;
    }
  }

  return 0;
}

(Inlined for posterity, since this comes from an uncommitted working directory.)