"check latitude of antimeridian crossing points" test failing on non-intel

[df7cb]

The "check latitude of antimeridian crossing points" test in h3_postgis/test/sql/postgis.sql is failing on all architectures except x68:

11:16:27 diff -U3 /tmp/autopkgtest.AAGzLw/tree/h3_postgis/test/expected/postgis.out /tmp/test-17/h3_postgis/test/results/postgis.out
11:16:27 --- /tmp/autopkgtest.AAGzLw/tree/h3_postgis/test/expected/postgis.out	2025-02-10 10:01:18.000000000 +0000
11:16:27 +++ /tmp/test-17/h3_postgis/test/results/postgis.out	2025-02-10 11:16:25.692876337 +0000
11:16:27 @@ -103,7 +103,7 @@
11:16:27      ) AS q1
11:16:27  ) AS q2
11:16:27  WHERE ABS(ABS(ST_X(p)) - 180) < :epsilon;
11:16:27 - t
11:16:27 + f

https://jengus.postgresql.org/job/h3-pg-binaries/20/

A common problem here is that x86 is still using i386's extended-precision 80-bit floating point arithmetic by default, so the results are more precise than they actually should be with 64-bit FP. This can be disabled by adding the -ffp-contract=off compiler flag. However, this doesn't work here (the test still passes on x86), possibly the same would have to be applied on the PostGIS side as well.

One fix might be to tweak the "epsilon" value used (but playing with it for a moment didn't work, I don't quite understand that test yet).

(I disabled that test for the Debian package build for now, so h3-pg is up-to-date on apt.postgresql.org and in Debian now.)

[zachasme]

I think lowering epsilon until the test pass is an acceptable solution. I didn't actually make that particular test, but I think the current value was chosen somewhat arbitrarily.

I would love to automate testing for this, but I don't believe I can choose architecture for GitHub runners. 🤔 I would be very grateful if you could help figure out many zeroes to remove such that the test pass on other architectures.