Consider:

shmem [qurx] ~ > perl -CO

use Encode qw/ _utf8_on /;
use Devel::Peek;

$s = " ";
_utf8_on( $s );
substr $s, 0, 1, "\x{b5}";
Dump $s;
print length $s, $/;
print "\$s: '$s'\n";
__END__
SV = PVMG(0xf64ff0) at 0xedd8c8
  REFCNT = 1
  FLAGS = (SMG,POK,pPOK,UTF8)
  IV = 0
  NV = 0
  PV = 0xee2620 "\302\265"\0 [UTF8 "\x{b5}"]
  CUR = 2
  LEN = 10
  MAGIC = 0xf66010
    MG_VIRTUAL = &PL_vtbl_utf8
    MG_TYPE = PERL_MAGIC_utf8(w)
    MG_LEN = -1
1
$s: 'µ'
[download]

So this has nothing to do with IO layers; sysread apparently does substr, which just does the right thing. Magic.

"\x{b5}" is an iso-8859-1 (a.k.a) latin1 char and a valid UTF8 codepoint, whose UTF-8 hex value is 0xC2B5 (0302 0265 as octal) - MICRO SIGN (µ).

print chr hex 'b5' eq "\x{00b5}";
__END__
1
[download]

Your file handle was set to raw, so bytes are read. Since perl places the byte into an UTF-8 string slot, it converts it from the internal representation into UTF8 and happily places its char hex value (2 bytes) into the PV slot, to satisfy the utf8-ness.

Similar to what happens here (reversed):

use Encode qw( from_to _utf8_on );
use Devel::Peek;

$s = "\x{b5}";
Dump $s;

from_to($s, 'latin1', 'utf8');
_utf8_on $s;
Dump $s;
__END__
SV = PV(0x234fa90) at 0x2375870
  REFCNT = 1
  FLAGS = (POK,IsCOW,pPOK)
  PV = 0x2372400 "\265"\0
  CUR = 1
  LEN = 10
  COW_REFCNT = 2
SV = PV(0x234fa90) at 0x2375870
  REFCNT = 1
  FLAGS = (POK,pPOK,UTF8)
  PV = 0x249b2c0 "\302\265"\0 [UTF8 "\x{b5}"]
  CUR = 2
  LEN = 10
[download]

Except that the 'magic' bits are missing (since none involved).