Confirmed that the problem also exists in Ubuntu 12.04.  

Confirmed that the break is in:

author	Steve Yin <steve_y@apache.org>	2013-11-29 13:03:27 (GMT)
committer	Caolán McNamara <caolanm@redhat.com>	2013-12-02 10:25:33 (GMT)
commit 76c549eb01dcb7b5bf28a271ce00e386f3d388ba

This is similar to a bug in Chrome.

KMFL (and other inputs methods) need to be able to delete text that has been previously entered. There are couple of ways to do this, one being to use a system call like "delete surrounding text". Another method is to simulate the backspace key. The former method is the preferred method because it gives better control over what is deleted.  However applications have to be written to support "surrounding text". Apps are supposed to report to ibus whether they support "surrounding text" calls and if they don't,  then kmfl falls back to simulating the backspace key. Unfortunately, Libreoffice 4.3 reports erroneously that it supports  "surrounding text" when in fact, it doesn't.

There is a work around. If you tell libreoffice to interface with ibus using xim rather than the "GTK ibus connector" it currently uses, then kmfl will delete characters using simulated backspaces. If you want to test it out to see if it works for you, try the following from a terminal:

XMODIFIERS=@im=ibus GTK_IM_MODULE=xim libreoffice --writer

If that works, then you can create a launcher for google-chrome with the following as the Command:

env XMODIFIERS=@im=ibus GTK_IM_MODULE=xim libreoffice --writer %U

calc and impress have similar issues. Note that writer 4.2.x did not have this issue with writer.

I'm pretty sure the bug is in the changes made to sw/source/core/access/accpara.cxx

Marking as confirmed based on Doug's comments, and on emails I have received since posting this bug.

No problem with Windows 2003, keyman and LO 4.3.3.2

The problem comes from the removal of this line:

    // If this object has the focus, then it is remembered by the map itself.
    nOldCaretPos = GetCaretPos();

SwAccessibleParagraph::SwAccessibleParagraph(
        SwAccessibleMap& rInitMap,
        const SwTxtFrm& rTxtFrm )
    : SwClient( const_cast<SwTxtNode*>(rTxtFrm.GetTxtNode()) ) // #i108125#
    , SwAccessibleContext( &rInitMap, AccessibleRole::PARAGRAPH, &rTxtFrm )
    , sDesc()
    , pPortionData( NULL )
    , pHyperTextData( NULL )
    , nOldCaretPos( -1 )
    , bIsHeading( sal_False )
    , aSelectionHelper( *this )
    , mpParaChangeTrackInfo( new SwParaChangeTrackingInfo( rTxtFrm ) ) // #i108125#
{
    SolarMutexGuard aGuard;

    bIsHeading = IsHeading();
    SetName( OUString() ); // set an empty accessibility name for paragraphs

    // If this object has the focus, then it is remembered by the map itself.
    nOldCaretPos = GetCaretPos();
}

Reviewed OpenOffice code as well.  The missing lines exist (but are commented out) in the current AOO code.  The commenting out occurred when Steven Ying implemented his huge "AOO IA2 enabled draft version 1" change on 27 Sep 2013.  https://github.com/apache/openoffice/commit/0deba7fbda3d9908785c25a443701a293b6f4e71#diff-50d752f41bb880abd1094d09d9e1a7fc

The current AOO comment is 
// If this object has the focus, then it is remembered by the map itself.
// not necessary to remember this pos here. Generally, the pos will be updated in invalidateXXX method, which may fire the
//Focus event based on the difference of new & old caret pos.
//nOldCaretPos = GetCaretPos();

Since this is in a call to the constructor, I see no reason not to re-implement this. There doesn't seem to be a good reason for why it was removed in the first place.

Confirmed that the bug also exists in Apache OpenOffice, and that uncommenting the one line fixes that one too.

Bug-fix submitted for review:  https://gerrit.libreoffice.org/#/c/12311

Justin Luth committed a patch related to this issue.
It has been pushed to "master":

http://cgit.freedesktop.org/libreoffice/core/commit/?id=817da76529aa39f641d76805d429b09681348811

fdo#85912 Delete surrounding text failing for input method calls regression.

It will be available in 4.4.0.

The patch should be included in the daily builds available at
http://dev-builds.libreoffice.org/daily/ in the next 24-48 hours. More
information about daily builds can be found at:
http://wiki.documentfoundation.org/Testing_Daily_Builds
Affected users are encouraged to test the fix and report feedback.

Cherry-picked for 4.3 branch
https://gerrit.libreoffice.org/#/c/12365/

accepted into the 4.3 branch.   Should be available with 4.3.5 in Dec 2014.

4.3 commit is here:
http://cgit.freedesktop.org/libreoffice/core/commit/?h=libreoffice-4-3&id=9a612f1d08eea858111ad5e161dc548c5173250a

(In reply to Commit Notification from comment #7)
> Justin Luth committed a patch related to this issue.
> It has been pushed to "master":
> 
> http://cgit.freedesktop.org/libreoffice/core/commit/
> ?id=817da76529aa39f641d76805d429b09681348811
> 
> fdo#85912 Delete surrounding text failing for input method calls regression.
> 
> It will be available in 4.4.0.
> 
> The patch should be included in the daily builds available at
> http://dev-builds.libreoffice.org/daily/ in the next 24-48 hours. More
> information about daily builds can be found at:
> http://wiki.documentfoundation.org/Testing_Daily_Builds
> Affected users are encouraged to test the fix and report feedback.

I have a pending Gerrit change essentially reverting that commit at https://gerrit.libreoffice.org/c/core/+/135868/1 because it causes issues with a11y and I'm unable the reproduce the original issue described here, s. commit message in that change for details. Any input appreciated.