Assuming relatively simple logic in reset and test functions (for example, no interactions between adjacent bits), I expected both to generate the same logic.
Understand that since the entire for loop executes in a single clock cycle, the synthesis expands it and generates a separate instance of test_array_bit for each input bit. Therefore, it is entirely possible that synthesis tools can generate identical logic for both versions - at least in this simple example.
And on this basis, I would (slightly) prefer the for ... loop version, because it localizes the program logic, while the "generating" version globalizes it, placing it outside the process template. If you find the loop version a little easier to read, you will agree at some level.
However, this does not mean that it is dogmatic of style, and your experiment illustrates this: loop synthesis for lower equipment. Synthesis tools are complex and imperfect pieces of software, such as highly optimizing compilers, and share many of the same problems. Sometimes they skip the โobviousโ optimization, and sometimes they perform complex optimization, which (for example, in software) is slower because its increased size broke the cache.
Therefore, it is preferable to write in the purest style where you can, but with some flexibility to work with the limitations of the tool and sometimes with real defects of the tool.
Different versions of tools remove (and sometimes introduce) such defects. You may find that the ISE option "use a new parser" (for parts up to Spartan-6) or Vivado or Synplicity gets this right where there is no ISE senior parser. (For example, signaling from procedures, older versions of ISE had serious errors).
It may be instructive to change the example and see if the synthesis can โget this rightโ (to produce the same equipment) for the simplest case and re-enter the complexity until it finds which design will fail.
If you find something specific in this way, itโs worth reporting it (answering your question). Xilinx used to encourage reporting such defects through its Webcase system; in the end they were even fixed! They seem to have stopped this, however, in the last year or two.