You can't depend on 
  <f si="..." t="shared"/>
 That is, it's _true_, but you can have a table with shared formulae
  that doesn't use it.  Compare M17:T28 (see below, uses shared) and
  C17:J28 (mostly no shared) in sample4

Looks like the result of a sweep-and-copy-{right,down} results in the
_new_ cells covered showing as 'shared':

         [ ][1][1][1][1]...
         [2][2][2][2][2]...
         [2][2][2][2][2]...
         ...

Presumably that one was right-then-down, down-then-right would give a
slightly different pattern
--------
Thinking about a pipeline...
 1) convert all variable references into (verbose!) elts:
     <!ELEMENT R EMPTY>
     <!ATTLIST R
        ac CDATA IMPLIED
        rc CDATA IMPLIED
        ar CDATA IMPLIED
        rr CDATA IMPLIED>

     where e.g. ac is 'absolute column'
    'D6' --> <R rc='D' rr='6'/>
     and
    '$E5' --> <R ac='E' rr='5'/>
--------
Identifying dates is . . . tedious.  They will be ints or floats (?),
with s="<int>", where the int is a 0-origin index into the list of
<xf...numFmtId="<bin>".../>
children of <cellXfs> in styles.xml, and bin is a built-in date format
code, see 18.8.30 numFmt (Number Format) in ISO/IEC 29500-1:2016(E) ==
C071691e.pdf
---------
Decided to distinguish between type (num, date, str, err, ...) and
class (cur, others to come?).  If non-standard code, just record that.