Merged as 84c906c and deployed; docs.bds.oncarlos.com is serving it. Landing recorded.
Review found one real defect and two smaller things, all fixed. The defect: the toolbar resolved a run carrying two font marks in the opposite order to both renderers — mono wins the cascade on screen and in the PDF, so a dual-marked run drew mono while the control read 'Sans serif' — and a test pinned that wrong answer, making the disagreement doctrine rather than an oversight. Only reachable through a hand-crafted update, which is exactly the case the surrounding comments claim to have considered. Also fixed: a guard that could not throw for any input, and a comment arguing the right decision from a backwards fact (the old default was serif, which carried no mark, so no pre-move document contains fontserif — what they carry is fontsans).
Recorded, not fixed: the three surfaces now agree on dual-mark order but only fonts.test.mjs pins it, so changing MARKS order or styleFor's switch would fail nothing. Closing that needs a fixture carrying a combination the editor cannot produce, and a cross-language render-and-diff — real scope, and the current behaviour is verified correct by hand against both renderers' source.
Screen and PDF defaults are now pinned against each other by a test that was verified by flipping the stylesheet back to serif and watching it fail. Page counts were measured before and after: four pages and two either way, nothing loosened.