Why are my icosahedron triangle subdivisions not equilateral (PostGIS)?

I am creating a geodesic polyhedron using PostGIS. I am dividing the base icosahedron points (see Appendix A), into the first level of Class I subdivisions.

The setup

For simplicity let’s take just the first triangle formed by these points:

CREATE TABLE pix (
    id bigserial
  , name text
  , geog geography(POINT,4326) -- use WGS84
);

insert into pix values (DEFAULT, 'China', st_point(122.3, 39.1));
insert into pix values (DEFAULT, 'Norway', st_point(10.53619898, 64.7));
insert into pix values (DEFAULT, 'Arabian sea', st_point(58.15770555, 10.44734504));

Which creates the base icosahedron triangle

The numbers of the triangle sides correspond to the row numbers shown in the table generated by this query

select p1.name
  , p2.name
  , st_distance(p1.geog, p2.geog) as dist
from pix p1
  cross join pix p2
where p1.id < p2.id

The points are roughly equidistant. There are small variances because the seed data above seems to have been generated assuming a sphere, not spheroid, which st_distance is returning – use_spheroid false in st_distance will yield much closer distances, but note that using false in all the presented queries does not change the problem described below.

First level of Class I subdivision

insert into pix (name, geog)
select p1.name || '-' || p2.name
  , st_project(p1.geog, st_distance(p1.geog, p2.geog) / 2, st_azimuth(p1.geog, p2.geog)) as geog
from pix p1
  cross join pix p2
where p1.id < p2.id

The above generates the mid-points of each of the pairs of points on the original triangle

And these have the following distances, generated by this query

select p1.name
  , p2.name
  , st_distance(p1.geog, p2.geog) as dist
from pix p1
  cross join pix p2
where p1.id < p2.id
  and p2.id > 3
order by dist

The problem

The distances of the segments 7, 8 and 9 (around 4000kms) are much longer than the distances of segments 1-6 (around 3500kms).

Why is this algorithm not creating something more close to equliateral triangles? Is my use of st_azimuth and st_project wrong? Or did I miss some more basic maths about projections of points onto spheres?