Why is one bothered about chain-stretch ? The greater the elongation of the chain, the more it over-loads individual teeth of your cassette and chain-rings - they can then be worn/damaged very fast. Worn chain-rings cause chain-suck, and worn cassette teeth cause chain-skipping ...and replacing them costs you a lot extra. These components should last through 3-to-6 chains, if the chains are maintained well and replaced timeously.
Norms when new : On new chains, the pitch for single links should be exactly 0.50" (ie link-pairs 1.00"). Thus 12 link-pairs on a new (or unstretched) chain, should measure exactly 12.00" = 304.8mm, but there are always manufacturing inaccuracies. On a range of brand new chains, I measured "stretch" to be in the range of -0.10% to +0.15% (measured using a long steel ruler over 48 link-pairs).
Replacement Criteria : For normal maintenance (chain-suck not an issue), chains should be replaced at about +0.5% overall "stretch". This +0.5% = 1/16" = 1.5mm over 12 link-pairs.
By +1.0% stretch (1/8" or 3.0mm over 12 link-pairs), chains will have done significant damage to both the chain-rings and rear sprockets. Then when you install a new (or low-stretch) chain, it will inevitably skip over the worn teeth on the rear sprockets during high-load pedalling ; this is very disconcerting and its impact loading will do further damage and will do it quickly.
If you have installed a new granny-ring or flipped the ring to use its unworn faces, for the purpose of alleviating chain-suck, then replace the chain preferably ; and definitely if its stretch is greater than +0.3%.
An alternative is to measure over 48 link-pairs. Remove the chain from the bike which is easy if you use an SRAM type "power-link" or any other of these universal type links ; clean the chain to remove any dirt packed into its joints which could affect the measurement ; hang the chain up with a small weight at the bottom ; use a long steel rule to measure the total pin-to-pin distance (see Tierney's photos - but over 48 link-pairs). The +0.5% stretch will now be 1/4" or 6.1mm which can be measured more easily and accurately over the longer chain length.
Fast and Simple Method : Another very simple method I use to assess chain wear is to have three chains always hanging next to each other (full-length) from the same thin steel rod in my garage ; they only needed to be measured the first time. One chain is badly worn at +1.0% wear, the next is normally worn at +0.5% wear, and the nearest is brand new. As soon as I've removed my current chain from the bike (in a jiffy with its "power-link"), I hang it on the same rod - with an instant view of its "stretch". Typically, chains on most bikes have a few more than 100 single links (50 link-pairs) ; when a chain has stretched 0.5%, its lower end will be about half a link longer than the new chain - then put the new chain on the bike, and buy a replacement new chain.
Measuring Gauges : Park Tools (and others) make gauges for measuring chain stretch. However, I have experienced very inaccurate results using it, when compared directly with the more accurate methods mentioned above. I also consider that Park's recommendation of a +1.0% stretch allowance is far too great (although it is noted that this type of gauge often reads +1.0% when the real stretch is actually +0.5%).
I also repeated my investigation into "stretch" of brand new chains on the same chains indicated in "Norms when new", using a borrowed Park Tool. This showed "stretch" of +0.20% to +0.40% (using a method to interpret the Park reading). So the Park Tool is conservative, but unneccesarily and wastefully so if the +0.5% criterion is used for discarding chains. Park's own criterion of +1.0% gets around this to some degree but leaves a wrong understanding of the issue. There are fundamental geometric reasons why this device (and other generics based on the same idea) will over-measure, and the degree of over-measurement gets worse as the chain wears. I don't know if any generics have some way of mitigating the problem and providing more accurate measurements, but it would appear not.