Thanks to @whuber for setting me on the right track here. Looks as if there will be no additional answers forthcoming, so will settle this question by posting my own observations that may be useful for others learning about distances, clustering, and projections.
The following R code, using the
sp packages demonstrates that careful selection of the right projection can give an accurate distance matrix (where accurate is defined as geodesic distance) when points are up to 2000 km apart (axes are in metres).
## Produce 200 randomly positioned geographic coordinates
## in central Canada
xyLatLon <- data.frame(lon=(runif(200)*-30)-85,
## Convert to a Lambert Conformal Conic projection that should
## reasonably approximate the true distance
newProj <- "+proj=lcc +lat_1=49 +lat_2=77 +lat_0=63.390675
+lon_0=-91.86666666666666 +x_0=6200000 +y_0=3000000
+ellps=GRS80 +units=m +no_defs"
xyLcc <- spTransform(SpatialPoints(xyLatLon, proj4string=CRS("+proj=longlat")), CRS(newProj))
## Find the geodesic distance matrix from geographic coordinates
## assuming the WGS84 ellipsoid
xyDist1 <- distm(xyLatLon, fun=distMeeus)
## Find the Euclidean distance matrix from the projection
xyDist2 <- as.matrix(dist(coordinates(xyLcc)))
## Find the Euclidean distance matrix of the geographic coordinates
xyDist3 <- as.matrix(dist(xyLatLon))
Plots of the elements of these three distance matrices are shown below. The plot on the left indicates that the projection selected is highly correlated with the geodesic distance across the range of distances used here. While the right plot demonstrates the considerable error that would be expected if unprojected geographic coordinates were to be used.