a STR is an unsigned long int.

One special STR value, NOSTR, is used to represent the zero-length
string.  No other collapsing of identical-valued strings is done.

The low 24 bits of a STR are an index into a table, allowing a little
over 16 million strings.  This table is kept in 4K-entry segments for
the sake of storage efficiency.  Each entry corresponds to one STRDESC,
but if the string is short enough (strlen(str) < sizeof(STRDESC)), the
bytes of the string are overlaid on the STRDESC.  Note that when the
string is overlaid this way, the trailing \0 is included, because
there's then no explicit length field available.  A vector of bits is
kept to flag when this is done.  A non-overlaid string has .u.taken
filled in, with .u.taken.core being meaningful only when .incore is
set.  .u.taken.usestamp is a last-use serial number, used to move
strings around to improve locality in the string file; .u.taken.fileoff
is an offset in bytes into the string file itself.  The trailing \0 may
not be stored explicitly; that's why there's a length field.  STRDESCs
not currently in use are chained together with the u.free.link field,
rooted at freedescs, have their .free bit set, and their STR values in
.u.free.id; others never have .free set.  Running out of STRDESC room
is a fatal error; I don't expect this to be a problem.

Space in the strings file is allocated in 4-byte units.  The file
length is known; free spaces in the file are kept in FREEFILE
structures.  When the array of FREEFILE structures fills up, the file
is compacted, moving all free space to the end and coalescing it into a
single block.