When does client-side routing still pay off now that browsers offer bfcache, prerendering and cross-document view transitions?

Open question

Client-side routing (intercepting link clicks, calling history.pushState, rendering the next view in place) was adopted to avoid full page loads: no blank flash, preserved in-memory state, instant transitions. Its costs are well documented: the server must serve the application shell for every route and report "not found" some other way; scroll position, focus and the document title must be restored by hand; analytics and error tracking need route hooks; and no navigation works until the JavaScript bundle has arrived and run.

Browsers have since added features that address the original motivations for multi-page sites. The back/forward cache restores a previously visited page from memory, provided nothing on the page (MDN names the unload handler) blocks it. The Speculation Rules API lets a page declare which documents should be prefetched or prerendered, so the next page can be ready before the click. Cross-document view transitions animate between two documents, which the MDN page notes was historically impossible.

Given these, for which kinds of site (content, dashboard, editor) and which interaction patterns (frequent navigation, long-lived widgets such as a media player, large in-memory state) does an in-page router still improve the measured experience, and where does a multi-page architecture with speculation rules and view transitions now match it at lower cost?

What a useful answer contains

The type of site, the navigation frequency and the state that must survive a navigation; measurements of navigation latency, LCP, INP and CLS for both architectures on the same content, with the browser versions and the share of visitors whose browser supports prerendering; the share of navigations that were prerendered or restored from bfcache; the cost side (bundle size, routes that broke on hard reload, how 404s are reported); accessibility findings (focus, announcements, reading position); and whether the team kept the architecture a year later. Answers from framework vendors should say so, and single-site anecdotes should be labelled as such.