This is the most successful relay design that I've come up with so far:
View from underneath the relay, showing the strut that the pole pivots on:It's an SPDT relay. The nail that is used as the core of the coil also serves as the normally open contact.
It's printed all in one piece - this seems preferable to assembling multiple pieces (as I did in an earlier version), because assembly brings with it the potential for imprecise alignment of pieces.
The moveable pole of the relay is mounted on a thin flexible strut. It pivots on this strut - flexing the strut - as the relay is activated and deactivated. In earlier designs I had this strut connected to the body of the relay at only one end, but found that this gave too many degrees of freedom (up/down, back/forth, pivot) and it became difficult to control the springiness of the pole.
An OpenSCAD file for this relay can be found here.
A dab of epoxy glue is used to keep the pole nail and the normally-closed nail firmly in place. The coil nail fits tightly enough in its holder that it doesn't require any glue - it can easily be removed for reuse in future design iterations.
The nails are galvanised clout nails with shaft diameter of 3mm and a head diameter of 9 or 10 mm. The heads are fairly uneven, and often the centre of the head is offset from the centre of the shaft. The coil and NC nails are 30mm long, the pole nail is 20mm long.
The coil is made from 1500 turns of 40swg enameled copper wire, it has a resistance of 35 ohms.
The operating voltage of this relay was found to be 3.1V, the release voltage 1.2V.
The resistance of the contacts is undesirably high - the normally closed contact has a resistance of 6.2 ohms, the normally open contact has a resistance of 1.2 ohms. (Measured with a 100 ohm load). Off-the-shelf general purpose relays typically have a contact resistance of less than 0.1 ohm.
The operate time of the relay is about 8ms, the release time about 9ms. The bounce time at the NO contact is typically < 1ms, the bounce time at the NC contact is much longer - the last bounce typically occurs at about 23ms after the removal of coil power.
I've operated this relay 60,000 times at 3Hz, switching low currents (about 1mA).