Text field performance

We are having some performance issues we think may be related to large
text field values, and many records. So, a couple questions...

1) Are text values always stored in the companion text file, or only if
they exceed a certain size? What size?

2) When the query executer reads tuples during a query, does it always
read the text companion file too? For example, the executer finds an
index match and then goes after the heap tuple to continue field
matching. Now, assuming the query does not contain any conditions
involving a text field, does the executer read the text field in even if
the tuple does not match and won't be included in the result set?

3) Is there any eqivelent to the CLUSTER functionality for text
companion files? Is there any such thing planned? What's the best way
to reduce fragmentation in a text companion file?

TIA

-Glen