The i18n self test says:

Specify that consumers of strings should use heuristics, preferably based on the Unicode Standard first-strong algorithm, to detect the base direction of a string except where metadata is provided. more (emphasis is mine, I.H.)

This requirement seems to be contradicted by the spec. At the moment, the EPUB 3.3 spec says

Inherent directionality specified using [Unicode] takes precedence over this attribute.

The way I can read this is that the first-strong algorithm in Unicode wins over the value of the dir attribute.

That would certainly be wrong; the Publication Manifest example shows the situation where the value of dir must be followed in contrast to the first strong algorithm.

Maybe that was not the intention in the EPUB spec, but then it needs clarification.