The way we do heuristic detection has the drawback that it may cause a reload in the mid-parse. Superficially, it seems that other browsers don't suffer from this problem. We should investigate what Trident and WebKit do about Japanese encoding detection and align our detection method with whichever makes more sense: the Trident method or the WebKit method. The WebKit code is http://trac.webkit.org/browser/trunk/Source/WebCore/loader/TextResourceDecoder.cpp#L157 We should also find out why Blink devs threw away the WebKit code and started using ICU's detector instead: https://mxr.mozilla.org/chromium/source/src/third_party/WebKit/Source/platform/text/TextEncodingDetector.cpp#40

The WebKit method would probably mean undoing bug 620106 and returning to a state similar to the one the HTML5 parser was in before bug 620106 was fixed. As is usual with bugs related to encoding detection, http://planet.debian.or.jp/ has fixed itself since.

FWIW, the site reported at https://webcompat.com/issues/15193 should render legibly when using the Japanese localization of Firefox. (But, yeah, having it be localization-dependent isn't ideal.)