0002 · The marked view may inject

Status: accepted, 2026-08-29. Depends on: 0001, which requires a second exception to no-injection to carry its own record. Issue: #6

---

The rule this bends

Decision 0001: a typed block lowers to the LaTeX its fields describe and nothing else. No support package, no wrapper environment, no helper command, however much the output would benefit.

xtex build --marked renders every revision visibly — inserted text coloured, deleted text struck through. There is no way to do that with kernel commands alone. It has to emit \textcolor and a strikethrough command, and it has to load the packages that define them.

Why it is permitted here

Because the property that no-injection protects does not apply to this view.

Property B says annotating must never change the rendered page. The whole purpose of --marked is to change the rendered page: a reviewer opens it precisely to see what an ordinary build hides. Applying the property here would be a category error, not a violation caught.

The rule and the property are not separable. No-injection exists because injection breaks property B and collides with the author’s packages and category codes. Where property B does not apply, the first reason is gone; the second is handled by the bounds below.

The bounds

1. --marked output is never the artefact of record. It is written to a distinct filename. A journal receives what xtex build produces, and that build is unchanged by this decision.
2. The normal build is byte-identical whether or not --marked was ever run. Testable directly, and it is the test that keeps this decision from leaking.
3. The injected set is fixed and listed here, not open to growth by whoever next needs something: \usepackage{xcolor} and \usepackage[normalem]{ulem}, plus \textcolor and \sout at each revision. Adding to that list is a change to this record.
4. A collision is the author’s to resolve, and it is visible. If the document already loads ulem without normalem, --marked may alter emphasis rendering. That is a defect in a review artefact, not in a submission, and it is why bound 1 exists.

What was rejected

Kernel-only marking. \underline exists; no kernel strikethrough does. Rendering deletions as [deleted: …] was considered and rejected: it makes a long deleted paragraph unreadable, which defeats the view’s only purpose.

Requiring the author to load the packages. It moves the injection into the source, where it would then appear in the submitted .tex — strictly worse. The build that must stay clean would become the one carrying review machinery.