Decisions that must not be undone

Both entries below describe code that existed, was measured, and was deleted. Both are the kind of thing a newcomer proposes in their first week.

The raster path was removed on purpose

convertDeckRaster / rasterizeSlide and the html2canvas dependency are gone, and nothing should reintroduce a rasterizer in src/.

A slide flattened into a full-bleed picture has no model to re-import: it is the one output that can never re-enter the loop, so it is a failure wearing a success's clothes. The cost is real and was accepted: hostile HTML no longer always yields some file. A warned failure beats a file that can never come back.

The obvious follow-up, "then render .pptx → PNG directly", was asked and answered: it cannot run in a browser. It needs a PowerPoint-grade renderer, and the slides that would most want a preview are the carried ones, which by definition have no model to draw from. The real options (PowerPoint COM Slide.Export, LibreOffice headless) are out-of-process and host-dependent, so a preview generator is a test utility, not a package feature. One under test/ for visual review of the corpus is welcome; nothing goes in src/.

Canvas use that remains in src/heuristic/ is not this: it rasterizes individual CSS gradients and re-encodes images, and each of those is a modeled element. It is not slide rasterization.

There is no raw OOXML work, and there must not be another repair layer

Everything goes through ts-pptx's DSL.

src/repair/repair.ts used to rewrite ppt/slides/slideN.xml, presentation.xml, [Content_Types].xml and the .rels parts after the writer had run. It was measured against ts-pptx 3.0.0 across the whole corpus and deleted. Four of its five rules never fired, and both that did were wrong: one flattened an inherited autofit into an explicit one, the other deleted every speaker note in the deck. It also swallowed exceptions and returned the unrepaired file, so its output depended on whether something threw.

The lesson is worth more than the code was: a layer that silently "fixes" things hides both its own obsolescence and its own damage. Every one of those defects was invisible from outside: the deck opened, the text was there, and only the notes were gone.

test/oracle/writer-output.test.ts now asserts the conditions the module existed for do not occur. If the writer regresses, the answer is an upstream issue and a failing test, not a second repair pass. If you find yourself reaching for a post-write rewrite, read that file first: it names each rule and why it went.