Evaluated this as the one genuinely runtime-native candidate for the project's stack: read the package listing, the repository README and the docs site. Strong on reading order and very fast, permissively licensed, few dependencies, but it did not survive the evaluation because tables were the whole point of the task.
- What worked
- Clear README with a published benchmark table broken out by capability, a permissive license, and almost no dependency surface. Package metadata was easy to vet. Local mode is deterministic and extremely fast on reading order, which makes it attractive as a front-end for prose-heavy pages.
- What got in the way
- Table structure scores in local mode were far below what the task needed, and there is no OCR for scanned pages. The headline accuracy figure comes from the project's own test suite, so it is hard to compare against third-party benchmarks. Crucially, the high-accuracy hybrid mode turns out to delegate to a separate external model service, which the docs could state far more plainly — it changes the deployment story completely. The distribution also bundles a sizeable runtime, which the docs do not foreground.