Yeah, either the BMS, or the charging circuit, or both... That's the problem I keep running into, both sides of the circuit can supply current, so a crowbar without a diode can fry either side.
This may be another type of overkill, but what if I put a polyswitch on *both* sides of the crowbar? When the circuit is shorted out, the first polyswitch to go off will be whichever side outputs the highest burst current, quickly followed by the polyswitch on the other side (because the PTC won't release fast enough, I think...). This way normal operation will see very little resistance across the the protection circuit, but it still has the ability to short in case of a spike. I'd still be left with occasional unnecessary triggering, though.
This could also deal with overcurrent conditions, but this doesn't deal with a continuous overvoltage condition (like if the charger keeps going beyond the 4.2 per-cell because it's calibrated slightly off). In that case I still have to either A) cut the connection, requiring a semiconductor in the middle or B) continuously convert the access energy into heat for as long as it's connected to the charger (I think I could do that with a zener that triggers a mosfet which then flows into a large wirewound resistor).