CONFIRMED for both 5.3.8 and 5.3.7 on Arch64, and for 5.3.4 on an unknown Linux.

Note however, that the test script provided by OP is wrong. It should be:
------------ BEGIN CODE ------------
$tmp = tempnam(sys_get_temp_dir(), 'test_');

$stream = fopen($tmp, 'w');
$data = "";

while(!feof($stream)){
  if(false === ($data = fread($stream, 8192))){
    break;         // ^---- no dot here
  };
}
------------- END CODE -------------
OP's code will fail regardless of the bug, because .= always produces a string.

This is not a bug. fread only returns false if given invalid arguments. The bug is that you try to read from a stream that's write-only.

This C program has analogous behavior:

#include<stdio.h>

int main(void)
{
    FILE *f = fopen("/tmp/foobaz", "w");
    printf("feof: %d\n", feof(f));
    printf("read: %zd\n", fread((char[100]) {}, 1, 100, f));
    printf("feof: %d\n", feof(f));
    return 0;
}

gcc --std=c99 h.c && ./a.out
feof: 0
read: 0
feof: 0

Let me change your code just a bit:
-------------------------------------------
#include<stdio.h>

int main(void)
{
    FILE *f = fopen("/tmp/foobaz", "w");
    printf("feof: %d\n", feof(f));
    printf("ferror: %d\n", ferror(f)); // <- HERE
    printf("read: %zd\n", fread((char[100]) {}, 1, 100, f));
    printf("feof: %d\n", feof(f));
    printf("ferror: %d\n", ferror(f)); // <- AND HERE
    return 0;
}

gcc --std=c99 h.c && ./a.out
feof: 0
ferror: 0
read: 0
feof: 0
ferror: 1

-------------------------------------------

In PHP there is no `ferror`. `feof` does work of both stdio's `feof` and `ferror`, as described in the documentation [1]: "Returns TRUE if the file pointer is at EOF or an error occurs".
Therefore, if we use analogy to stdio, `feof` should return `TRUE` in this case.

[1] http://pl.php.net/manual/en/function.feof.php

I feel I should say that just because the behaviour is analogous to the way the underlying C works does not make it correct nor useful.
While I concur that one should not really attempt to read from a write-only stream, I DO feel that any attempt to perform such an action should warrant an exception/warning/error of some sort. This is currently not the case and as a result can lead to potential indefinite loops given recommended usage in the PHP documentation.
Perhaps this should be a feature/change request?

You are right: no one should try to read from write-only stream. Also no one should try to divide by 0 and no one should pass something other than stream to `fread`. But people do it all the time and that's why errors, warnings, exceptions and status codes exist. Using an invalid type of a stream is an error too and should be reported.

If returning `FALSE` from `fread` breaks something, than I agree: `fread` bahaviour should not be changed. However a warning does not break anything, so it can be emitted. Since `php_stream` stores `fopen` flags in `mode`, I believe this can be done easily:
--------------- BEGIN DIFF ---------------
diff php5.3-201201041830/ext/standard/file.c php5.3-201201041830-modified/ext/standard/file.c
1893a1894,1898
> 	
> 	if (strpbrk(stream->mode, "+r") == NULL) { /* r or any + mode is fine */
> 	    php_error_docref(NULL TSRMLS_CC, E_WARNING,
> 	          "Reading from a write-only stream");
> 	}
---------------- END DIFF ----------------
After this modification some of ext/standard/tests/file/007* tests will fail, but this is expected.

I agree that this case is a feature request.

All right, I'll give a bit more thought to this.

The problem here is more general, as there is no way to detect file access errors because, as already stated, the ferror() function of the libc has no a corresponding implementation under PHP. In my opinion, on an application language as PHP is, such a low level function should not exist and instead should be fread() to return FALSE when it fails reading from a file because a I/O error of any kind happened, and should emit E_WARNING with appropriate message. The same should apply to any other file access function as well.

If it may be useful, i wrote this script that simulates a I/O error on Linux, and illustrates how the access errors cannot be detected from PHP code:

<?php
// fread() fails to detect I/O error, bug #60671.
// Test code working on Linux only.

//
//  Set a safe environment:
//

error_reporting(-1);
ini_set("track_errors", "1");


function my_error_handler($errno, $message /*, $filename, $lineno, $context */)
{
	throw new ErrorException("$errno: $message"); 
}

set_error_handler("my_error_handler");


function main()
{
	// Simulate I/O error in Linux (credits http://unix.stackexchange.com/a/6302):
	// reading of the first 512-bytes block of /proc/self/mem causes EIO in libc.
	// $ cat /proc/self/mem
	// cat: /proc/self/mem: Input/output error
	$r = fopen("/proc/self/mem", "r");
	while( ! feof($r) ){
		$s = fread($r, 100);
		echo "read: "; var_dump($s);
	}
	fclose($r);
	// Outcome: no errors whatsoever, only displays:
	// read: string(0) ""
}

main();

This issue is resolved for all actively supported PHP versions[1].
fread() returns false for write only streams.

[1] <https://www.php.net/supported-versions.php>