I don't know whether this bug is a fault of LR or PS, so I tagged this problem with both products.
LR 3.4.1 and PS CS4, Windows XP 32 Bit SP3.
Steps to reproduce:
1. Select a RAW in develop module of LR
2. Delete all local corrections and spot removals
3. Draw some adjustment brush strokes
4. *** save metadata (Ctrl-S) ***
5. Delete all brush strokes
6. Open as Smart Object in PS, or alternativley Open in PS, with PS doing the rendering in PS-ACR (the bug does *not* happen when LR is rendering the image!)
Result: In PS, the deleted adjustment brush strokes reappear!
The same happens a) for a virtual copy made after step 5, and b) even after modifying some other settings, e.g. contrast.
And now for the funny thing: Adding a new different brush stroke will cause the image to be rendered *correclty* in a subsequent "open as smart object" or "edit in PS" (i.e. only the new, different brush stroke appears). *Deleting* all brush strokes will again make the old brush strokes reappear!
Note: As one would expect, the old brush strokes disappear in a subsequent "open as smart object" or "edit in PS" after the metadata are again saved to XMP. The same can be achieved by deleting the .XMP sidecar in case of a proprietary RAW file.
Theory: When a RAW ist passed from LR to PS together with the development settings to be rendered in Photoshop's ACR, the passed settings are somehow *mixed* with the settings stored in the XMP metadata of the RAW file (i.e. in its sidecar for proprietrary RAWs). When no brush strokes are present in the passed settings, PS-ACR will take the brush strokes from the XMP. This is of course wrong, because the XMP metadata may be outdated, or in case of a virtual copy may be even totally different. The XMP metadata has to be completely ignored by PS-ACR.
Note: This was also reported to be happening with PS CS5 as well. So it could be that the actual bug is not on the side of Photoshop/ACR, but already in LR where the settings get mixed. However, it could be photoshop, too.
Edit: Spot removal and graduated filter are affected in the same way (independently of one another). I didn't test the red eye correction.