I don't think that's the problem, actually the ancestor returned by rebasemerge
is the correct one (the first). The line with my... is returned by filemerge
that seems to override the modified ancestor with a wrong one (it's the real
ancestor and we don't need it).

Initial situation:

@  3:66559d44c7db3a5155bbaf8c7a364504a42932e2 D                               
|                                                                             
| o  2:7e1afe9214b270fa61150d316a28758ca92f956d C                             
| |                                                                           
| o  1:0a6620c3c26a757c498308a9f6773d154681bde0 B                             
|/                                                                            
o  0:1e635d440a7348d13e5fafba3efc35d9a54c2cb1 A                               

rebase onto 3 starting from 1                                                 
merge against 1:0a6620c3c26a
first revision, do not change ancestor
resolving manifests
 ancestor 1e635d440a73 local 66559d44c7db+ remote 0a6620c3c26a
picked tool 'internal:merge' for a (binary False symlink False)
my a@66559d44c7db+ other a@0a6620c3c26a ancestor a@1e635d440a73

Intermediate situation:
o  4:0b69665d48b4b5cacd2eb0e92d8cfa5a6a1d2f16 B                               
|                                                                             
@  3:66559d44c7db3a5155bbaf8c7a364504a42932e2 D                               
|                                                                             
| o  2:7e1afe9214b270fa61150d316a28758ca92f956d C                             
| |                                                                           
| o  1:0a6620c3c26a757c498308a9f6773d154681bde0 B                             
|/                                                                            
o  0:1e635d440a7348d13e5fafba3efc35d9a54c2cb1 A                               
                                                                              
merge against 2:7e1afe9214b2
Invoked newancestor
Using 0a6620c3c26a757c498308a9f6773d154681bde0 instead of
1e635d440a7348d13e5fafba3efc35d9a54c2cb1
resolving manifests
 ancestor 0a6620c3c26a local 0b69665d48b4+ remote 7e1afe9214b2
picked tool 'internal:merge' for a (binary False symlink False)
my a@0b69665d48b4+ other a@7e1afe9214b2 ancestor a@1e635d440a73

The line "Invoked newancestor" shows which revision has been returned by
rebasemerge.
I think we need to find a way to prevent filemerge from using the original ancestor.