Kernel 2.6.31.x

dmesg:
Synaptics Touchpad, model: 1, fw: 6.3, id: 0x1c0b1, caps: 0xa04751/0x0

but /proc/bus/input/devices:

I: Bus=0011 Vendor=0002 Product=0007 Version=01b1
N: Name="SynPS/2 Synaptics TouchPad"
P: Phys=isa0060/serio1/input0
S: Sysfs=/devices/platform/i8042/serio1/input/input15
U: Uniq=
H: Handlers=mouse1 event2
B: EV=b
B: KEY=420 70000 0 0 0 0
B: ABS=11000003

So even when capabilities clear say only 2 buttons, driver claims there 
are 3 of them.

The reason most likely is initialization sequence. 
psmouse_switch_protocol() unconditionally sets supported buttons:

        input_dev->evbit[0] = BIT_MASK(EV_KEY) | BIT_MASK(EV_REL);
        input_dev->keybit[BIT_WORD(BTN_MOUSE)] = BIT_MASK(BTN_LEFT) |
                BIT_MASK(BTN_MIDDLE) | BIT_MASK(BTN_RIGHT);
        input_dev->relbit[0] = BIT_MASK(REL_X) | BIT_MASK(REL_Y);

before even starting hardware detection and knowing real capabilities. 
Detection for specific hardware won't change it (possibly only extend).

Is it OK to move button bits setting into ps2bare_detect()? This seems 
to agree with comments in psmouse_extensions() as well:

/*
 * Okay, all failed, we have a standard mouse here. The number of the 
buttons
 * is still a question, though. We assume 3.
 */

All other detection routines seem to be setting those bits already.