potential/backend: anchor scalar coordinates in migrated analytic/disk compute methods

[jobovy]

What

Anchor scalar/numpy coordinate operands onto the active backend at the start of the migrated compute methods of 5 potentials (SpiralArmsPotential, EllipticalDiskPotential, KuijkenDubinskiDiskExpansionPotential, PowerSphericalPotential, PowerSphericalPotentialwCutoff), via the established per-method coerce_coords convention (the same one EllipsoidalPotential._anchor_phi uses).

These methods feed coordinates straight to xp.<fn> (cos/sin/exp/sqrt/log/abs/sign), which torch rejects on a python-float / numpy.float64 / numpy.ndarray. #990's potential_physical_input boundary coercion covers the public array-evaluation path for backend-compatible potentials, but it is bypassed by:

• direct private-method calls (the test_2ndDeriv / test_forceAsDeriv numerical-derivative path through undecorated mock wrappers),
• the internal axisymmetric phi = 0 reset (mockFlatEllipticalDisk),
• parameter/t forwarding (kuijkendubinski_t_forwarding).

Result — 6 genuine torch reds flipped

test_2ndDeriv_potential[CorotatingRotation3DSpiralArmsPotential]
test_forceAsDeriv_potential[CorotatingRotation3DSpiralArmsPotential]
test_2ndDeriv_potential[mockFlatEllipticalDiskPotential]
test_amp_mult_divide[mockFlatEllipticalDiskPotential]
test_forceAsDeriv_potential[mockFlatEllipticalDiskPotential]
test_kuijkendubinski_t_forwarding

(verified fail-without / pass-with under forced torch, regen mode.)

PowerSpherical / PowerSphericalwCutoff flip no currently-red entry — #990's boundary masks their direct-compute-method gap — but they were migrated before this convention existed, so they're brought to per-method coerce_coords parity (byte-identical; closes the latent direct-call torch gap). Honest defense-in-depth, not a red-flip.

Safety (adversarially reviewed — two independent passes)

• numpy byte-identical — coerce_coords is an object-identical pass-through when xp is numpy; verified identical SHA-256 output hashes across all 5 families (incl. the KuijkenDubinski lambda→def closure conversion, which is behaviorally identical).
• jax value-identical — worst diff 3.95e-14 across the 5 families.
• Coverage — the new coerce_coords lines execute on the numpy test_potential path (build.yml numpy jobs, coverage-uploaded); no backend-only line.
• Altitude — per-method is correct: the bypass is the mock/direct-private layer, not a galpy mixin; there is no shared choke point for these 5 potentials.
• No ledger change (the 6 become XPASS, green via strict=False; pruned in the separate ledger-regen PR).

File-disjoint from #991 (_coerce.py/_namespaces.py). Base is current feat/backends (#990).

Out of scope (separate follow-ups the review surfaced)

• expwholeDiskMultipoleExpansionPotential 2ndDeriv — a construction-time numpy.ndarray * Tensor in MultipoleExpansionPotential.from_density (Finalize pvz etc. functions #75-adjacent).
• specialSpiralArmsPotential 2ndDeriv — a torch finite-difference precision mismatch, not a scalar crash.
• The larger _backend_compatible mock-bypass cluster (DehnenBar / Ferrers / SoftenedNeedleBar / …) — same vector, each needs its own per-method audit.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.93%. Comparing base (38dad76) to head (e0532c0).

Additional details and impacted files

@@              Coverage Diff               @@
##           feat/backends     #992   +/-   ##
==============================================
  Coverage          99.93%   99.93%           
==============================================
  Files                254      254           
  Lines              39839    39897   +58     
  Branches             839      839           
==============================================
+ Hits               39813    39871   +58     
  Misses                26       26           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.