The answer to this one is semantics-squared… woohoo! This is as close as I intend to get to pedantry, and figure that at the end of the day, the approach I *don’t* use is no better or worse. (Both approaches are ugly hacks in comparison to the ideal, if one considers things with the attitude of a purist.)
Desire for small, shiny objects aside, the primary content of the heading is definitely the copy, *not* its bitmapped representation… the bitmap is the fallback at the _semantic_ level. Thus the bitmap gets relegated to the stylesheet, while the actual heading _copy_ finds its way into the element’s textnode (rather than as some alternate value).
In a perfect world, the handheld media type would be respected, and the screen media type ignored by handhelds per the spec (7.3), unless the handheld user explicitly applies the screen stylesheet instead:
bq. Media types are mutually exclusive in the sense that a user agent can only support one media type when rendering a document. However, user agents may use different media types on different canvases. For example, a document may (simultaneously) be shown in ‘screen’ mode on one canvas and ‘print’ mode on another canvas.
So I wait for that Radiant Future, and implement a solution that looks forward (by keeping textual content at front-center) while managing to please most of the people most of the time.
On the other hand, use of RE-driven replacement validates the inline-image approach in terms of future-friendliness.
Bear in mind that in cases where the image *is* indisputably the primary content, this entire rationale gets thrown out on its ear.