The stereowiden filter uses a buffer, s->buffer[], and a pointer
within the buffer, s->write, to implement inter-channel delays.
The loop which applies the delayed samples turns out to be faulty.

   109      for (n = 0; n < in->nb_samples; n++, src += 2, dst += 2) {
   110          const float left = src[0], right = src[1];
   111          float *read = s->write + 2;
   112
   113          if (read > s->buffer + s->length)
   114              read = s->buffer;
   115
   116          dst[0] = drymix * left - crossfeed * right - feedback * read[1];
   117          dst[1] = drymix * right - crossfeed * left - feedback * read[0];
   118
   119          s->write[0] = left;
   120          s->write[1] = right;
   121
   122          if (s->write == s->buffer + s->length)
   123              s->write = s->buffer;
   124          else
   125              s->write += 2;
   126      }

For one, the buffer gets written past its end in lines 119-120, before
the bound check is done in lines 122-123.  This can be easily confirmed
by valgrind.

==3544== Invalid read of size 4
==3544==    at 0x593B41: filter_frame (af_stereowiden.c:116)
==3544==  Address 0xb1b03c4 is 4 bytes after a block of size 7,680 alloc'd
==3544==
==3544== Invalid read of size 4
==3544==    at 0x593B66: filter_frame (af_stereowiden.c:117)
==3544==  Address 0xb1b03c0 is 0 bytes after a block of size 7,680 alloc'd
==3544==
==3544== Invalid write of size 4
==3544==    at 0x593B79: filter_frame (af_stereowiden.c:119)
==3544==  Address 0xb1b03c0 is 0 bytes after a block of size 7,680 alloc'd
==3544==
==3544== Invalid write of size 4
==3544==    at 0x593B7D: filter_frame (af_stereowiden.c:120)
==3544==  Address 0xb1b03c4 is 4 bytes after a block of size 7,680 alloc'd

Also, using two separate pointers, s->write and read = s->write + 2,
does not seem to be well thought out.  To apply the delay of s->buffer[],
it is enough to read the delayed samples at the current position within
the buffer, and then to store new samples at the same current position.
Thus the application of delayed samples can probably be best described
with a single pointer s->cur.

I also introduce a minor change to ensure that the size of s->buffer[]
is always a multiple of 2.  Since the delay parameter is a float, it is
otherwise possible to trick the code into allocating off-by-one buffer.