Add reaction-based spherical-shell geoid projection

[gthyagi]

Summary

Supersedes #646 with a single, narrowly scoped commit from current
development (9da04b72). The original PR accumulated unrelated mantle
convection development; this replacement contains only the outstanding
spherical-shell reaction-projection changes.

The generic Stokes.boundary_normal_traction_integral() implementation is
already upstream through #648. This PR reuses that API rather than adding a
second implementation or changing shared solver/Cython code.

• Add projection="reaction" to the spherical-shell rotated-Stokes adapter.
• Normalize the reaction contraction by the matching finite-element boundary
  harmonic norm.
• Add serial and MPI coverage and document the alternative projection.
• Retain projection="centroid" as the compatibility default. Deprecating the
  ConvexHull workflow discussed in geoid harmonic projection re-derives the surface mesh with ConvexHull instead of integrating over the boundary facets we already have #647 remains a separate API decision.

No SUPG/SLCN, checkpoint, evaluator, scalar heat-flux, or solver-configuration
changes are included. The diff is five files: one postprocessing module, two
test files, the subsystem documentation, and the changelog.

Why and How

A nodal reaction represents an integrated force, not a pointwise traction.
When only a harmonic coefficient is needed, recovering pointwise P2 values and
then reconstructing a surface triangulation is unnecessary. It also encounters
the curved-boundary P2 issues discussed in #414, #431, and #633/#649.

For the scalar normal-reaction load R and nodal samples of the harmonic
phi = P_l(cos(theta)), the existing upstream integral API evaluates:

I(phi) = phi^T R - (1^T R / A) * integral_boundary(phi)

The adapter divides by BdIntegral(phi**2) and applies its existing boundary
sign and buoyancy scale. Numerator and denominator therefore use the same
faceted finite-element boundary geometry. Reaction ownership and reductions
remain the responsibility of the upstream API.

The reaction path needs no global boundary-sample gather, rank-zero
ConvexHull, or pointwise traction recovery. Consumers needing a nodal
topography field continue to use the existing pointwise APIs.

Fresh Validation of This Branch

Tested commit: fd0a3bcc. A wheel was built from this clean branch and installed
into an isolated target using the existing Pixi dependencies. PYTHONPATH
selected that installation for the tests; its imported UW3 had the upstream
reaction API and no AdvDiffusionSUPG class. The installed geoid.py checksum
matches the branch source.

Environment: macOS/arm64, Python 3.12, PETSc 3.25.0, Open MPI 5.0.10,
MPI-enabled h5py.

Test	Result	Time
test_1070_postprocessing_geoid.py	13 passed, including explicit reaction projection and invalid-selector coverage	Included below
test_1120_SLVectorCartesian.py	All 3 cases passed, including the case previously flagged in #646	27.11 s for both serial files
parallel/test_1071_spherical_shell_geoid_parallel.py	Passed on all 4 MPI ranks; eight response coefficients checked against Zhong Table 2	Approximately 5.18 s

The MPI comparison allows near-machine-precision reduction differences
(rtol=1e-13, atol=1e-14) rather than requiring bitwise equality.
Serial output contained three non-interactive plotting warnings and no failures.

Separately, the synchronized mantle integration branch passed strict
Level 1 selection: 1676 passed, 36 skipped, 2 expected failures in 1053.89 s.
That broader check is not presented as a full Level 1 run of this standalone
branch. Linux CI remains to be verified.

Build, extraction, and validation report.

Historical Production Evidence

These are the controlled Gadi results already reported in #646, not new
measurements of this extracted commit. The layered Zhong case used degree 5,
load depth 0.5, cellsize=1/64, a fitted 10^4 viscosity lid, P2P1, rotated
free slip, 192 ranks, and Stokes tolerance 1e-5.

Self-gravity response	Centroid error	Reaction error
Surface topography	-2.62%	+0.08%
CMB topography	+0.91%	<0.01% in magnitude
Surface geoid	-4.01%	+0.12%
CMB geoid	+1.19%	<0.01% in magnitude

Stage	Centroid	Reaction
Stokes solve	784.55 s	772.01 s
Response integrals, topography, and geoid	120.14 s	36.92 s

The response stage was 3.25 times faster. It includes common velocity/divergence
integrals, not geoid evaluation alone. The small Stokes timing change is
run-to-run variation; no Stokes speedup or whole-job memory reduction is claimed.
Velocity is unchanged by projection, including the approximately 8.46% residual
error in this case's small surface velocity.

Production benchmark
and historical results.

Review

@lmoresi: this isolates the spherical adapter from the unrelated changes you
flagged in #646. Please review its use of the already-merged reaction integral
and matching boundary norm. No generic reaction-recovery implementation is
being replaced. The mantle branch remains the development branch and is not
deleted when the superseded PR is closed.