> 
> AFAIK, even the mighty Apache does not have a mechanism for  
> automatically marking translations stale when you update the master.

The scoring should be done on the basis that the translations are not as
well maintained, so it shouldn't matter whether they are currently stale,
just that they are potentially stale.  (They should also be downrated
on the basis that they are not the official text.)

> We've had language negotiation for years and there's no "proper  
> implementation" in sight. Shouldn't this indicate that this instance  
> of architecture astronautics does not work or the cost of making it  
> work outweighs the benefit?

I'd say the real reasons are:

- they are not presentational, and in spite of attempts here to encourage
  a semantic web, most people commissioning web sites really want a page
  description language;

- they require server configuration, which is often reserved to the ISP
  (for cheap accounts) or the IT department, so difficult for designers 
  to learn and awkward for them to employ - although many technologies are
  popularly confused with HTML, HTTP tends not to be;

- being server features the configuration is not portable between servers
  (e.g. a move to IIS requires a specific directory layout).