Init TotalVerticalTransportPseudoVelocity for testing.

[overfelt]

Since tendencies now use TotalVerticalTransportPseudoVelocity, initialize to the same values as VerticalPseudoVelocity and TotalVerticalPseudoVelocity. Fixes #573

Do a running total of errors instead of zeroing out each time.

Checklist

• Building

  • CMake build does not produce any new warnings from changes in this PR

• Testing

  Perlmutter

  • CTests Pass
  • Polaris omega_pr Pass

[xylar]

Thanks @overfelt. Taking a look now.

[xylar]

Testing

This fixes the inf in FCTHNewInv and makes the FCT checks fail CTest, but two TEND tests now fail on Chrysalis. I ran CTests with Intel (Release) on the PR base c1aacdc2d7 and on this branch, with #574 merged into both so they configure (#572), and the 260911 sphere mesh. I also ran this branch with the new TotalVerticalTransportPseudoVelocity init line removed, to check that the inf is now caught:

	CTest	FCTHNewInv
base	50/50 pass	inf on plane, single-precision plane and sphere
this PR	48/50 pass	finite
this PR without the init line	47/50 pass	inf, all 3 TEND tests fail

The two remaining failures:

• TEND_SPHERE_TEST: FCTHNewInv LInf FAIL, expected 3.0541724683419424e-05, got 3.0543973828844884e-05. That expected value was never testable while the result was inf; the new value matches ExpectedFCTHInv to 9 digits.
• TEND_PLANE_SINGLE_PRECISION_TEST: FCTHProv, FCTHProvInv and FCTHNewInv (expected 0, got 1.5e-05), FCTHighAndLowOrderFlux_High and FCTAccumulateHighOrderFlux_0. These are the single-precision failures from FCT tendency tests swallow failures; FCTHNewInv is inf #573, unchanged from the base. All five compare near-zero expected errors with an ATol of 1e-10, which single precision cannot meet.

Work directory: /lcrc/group/e3sm/ac.xylar/polaris_1.1/chrysalis/test_20260928/omega575

---

Posted by Claude Code on @xylar's behalf. The testing, analysis and wording above are AI-authored; please check them accordingly.

[xylar]

Approved based on a look at the code changes and my testing with Claude.

Whoops, I jumped the gun. Need to take a more careful look.

[xylar]

@overfelt, I approved too soon. (Multitasking gone awry.)

Would you be able to look at the new failing tests on Chrysalis? I don't know my way around the CTests you added well enough to understand the failures yet.

I can give it a go if you can't.

[xylar]

Both remaining failures from my testing comment look like test-constant fixes. I haven't rebuilt with these changes.

For TEND_SPHERE_TEST, the LInf in the spherical ExpectedFCTHNew (line 222) needs to become 3.0543973828844884e-05, the value from my run; the L2 already passes. The old value predates the regenerated sphere mesh: 1a564eb updated ExpectedFCTHProv and ExpectedFCTHInv, but FCTHNewInv was inf at the time.

---

Posted by Claude Code on @xylar's behalf. The testing, analysis and wording above are AI-authored; please check them accordingly.

[xylar]

The single-precision failures are round-off, up to 1.5e-5 against expected values of 0 or ~1e-16. With RTol scaling those tiny values, this ATol sets the tolerance, and 1e-10 is too tight for floats. Making it precision-dependent like RTol at line 1979 would cover all five on Chrysalis:

const Real ATol = sizeof(Real) == 4 ? 1e-4 : 1e-10;

That is needed here and at lines 1495, 1541 and 1672, or once at the top of the function.

[overfelt]

I tried to test in Chrysalis but can't access the platform and assumed it was down for maintenance. Maybe something wrong with my account that I need to fix.

[xylar]

Chrysalis isn't down so indeed it seems like you need to check with @rljacob about getting your account reactivated once he's back.