potential/backend: anchor construction-time scalar phi to float64 in EllipsoidalPotential

[jobovy]

What

Anchor a construction-time python-scalar phi to float64 in EllipsoidalPotential._anchor_phi.

During __init__, normalize() calls Rforce(1., 0.) with plain python scalars before _backend_compatible is set, so the @potential_physical_input boundary does not coerce them. Under a forced non-numpy backend _anchor_phi then saw a dtype-less scalar R and fell back to xp.asarray(phi, dtype=None) — which on torch (whose CI jobs keep the float32 default, unlike jax's x64) produced a float32 phi, dropping the computed _amp to float32 and silently losing precision in every later float64-coordinate evaluation.

The fix: when R carries no dtype, anchor phi to xp.float64 (galpy's interior precision).

 if xp is numpy or is_backend_array(phi):
     return phi
-return xp.asarray(phi, dtype=getattr(R, "dtype", None))
+dtype = getattr(R, "dtype", None)
+if dtype is None:
+    dtype = xp.float64
+return xp.asarray(phi, dtype=dtype)

Result

Fixes the 6 oblate/prolate/triaxial Hernquist & Jaffe test_amp_mult_divide torch failures (baseline residuals ~1e-7, i.e. float32-magnitude, vs the 1e-10 tolerance). They flip xfail→XPASS (ledger stays green via strict=False; the stale entries are pruned in the separate ledger-regen PR, not here).

Safety (adversarially reviewed, CPU-forced)

• numpy byte-identical — the xp is numpy guard short-circuits before any new code; verified max abs diff 0.0 over a full evaluate/forces/2nd-deriv grid.
• exit-cast policy preserved — a genuine float32 R array still anchors phi to float32 (the dtype is None branch is only taken for a dtype-less scalar).
• autodiff w.r.t. phi untouched — a real backend phi array returns via the is_backend_array short-circuit (no cast/detach).
• jax unaffected (x64 already gives float64); no regressions across numpy/jax/torch in test_backend_ellipsoidal.py (687 passed).
• New test test_construct_normalize_amp_is_float64 reproduces the CI torch float32-default condition and pins the construction-time _amp to float64 (fails on baseline, passes with the fix). It also covers the new branch under the coverage-uploading test_backend* shard.

Notes / divergences

• Investigated under "stabilize the GL sum order", but the Gauss-Legendre quadrature was provably not the cause (vectorizing _potInt left the tests failing). The fix is a one-spot dtype anchor — the branch was renamed off gl-stabilize to reflect that.
• Follow-up (separate PR): the Ellipsoidal subclasses (PerfectEllipsoid, TwoPowerTriaxial/TriaxialHernquist/TriaxialJaffe, TriaxialGaussian, PowerTriaxial) still set _backend_compatible=True after normalize(), unlike the ~10 potentials backend: central coordinate coercion — fix systemic torch scalar-input rejection #960 reordered (only TriaxialNFW was reordered). Reordering them is the root-cause cleanup; this guard is the correct, byte-identical, independently-defensible boundary fix and keeps the branch covered.
• triaxialLogarithmicHaloPotential / mockRotatedAndTiltedTriaxialLogHaloPotential torch entries are out of scope — LogarithmicHaloPotential is not an EllipsoidalPotential subclass (already backend: central coordinate coercion — fix systemic torch scalar-input rejection #960-reordered; any residual float32 there is a different mechanism).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.93%. Comparing base (69659c6) to head (4effc36).
⚠️ Report is 2 commits behind head on feat/backends.

Additional details and impacted files

@@               Coverage Diff               @@
##           feat/backends     #989    +/-   ##
===============================================
  Coverage          99.93%   99.93%            
===============================================
  Files                254      254            
  Lines              39820    40276   +456     
  Branches             837      837            
===============================================
+ Hits               39795    40251   +456     
  Misses                25       25            

☔ 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.

[github-actions]

All-backend test status (jax / torch)

Commit a771a5bed478b5db2c9da183d394492dec6a7114

Green is achieved via the checked-in xfail-ledger (tests/backend_xfail.txt, applied xfail(strict=False)), so the metric to watch is the shrinking xfail count (burndown), not a raw pass count. A FAIL/ERR is an un-ledgered regression (reds the run). Because the ledger is non-strict, a now-passing ledgered test is a plain pass here (no per-push XPASS); burndown candidates -- in both directions -- are surfaced by the scheduled regen run, which rewrites the ledger from real outcomes. deferred is a separate burndown: tests skipped because they are unrunnable under the backend until the port is vectorized (see tests/backend_slow_skip.txt), e.g. the jax spherical-DF sampling/quadrature tests pending the Track F DF migration.

Overall: jax: 1056 passed · 266 xfail · 725 deferred | torch: 833 passed · 1210 xfail · 1 deferred

Ledger size: 2357 entries (jax=284, torch=2073).

Test shard	jax	torch
actionAngle	✅ 112 pass · 89 xfail	✅ 26 pass · 175 xfail
sphericaldf	✅ 164 pass · 26 xfail · 28 deferred	✅ 8 pass · 210 xfail
conversion + util + misc	✅ 85 pass · 6 xfail · 1 deferred	✅ 41 pass · 51 xfail
potential + scf + multipole	— (no result)	— (no result)
quantity + coords	✅ 287 pass · 49 xfail	✅ 189 pass · 147 xfail
orbit (energy/Jacobi + from_name)	✅ 0 pass · 0 xfail · 115 deferred	✅ 63 pass · 52 xfail
orbit + orbits (main)	✅ 0 pass · 0 xfail · 578 deferred	✅ 248 pass · 327 xfail
evolveddiskdf	✅ 35 pass · 0 xfail	✅ 32 pass · 3 xfail
jeans + dynamfric	✅ 17 pass · 2 xfail · 2 deferred	✅ 7 pass · 13 xfail · 1 deferred
qdf + pv2qdf + streamgapdf_impulse + noninertial	✅ 57 pass · 75 xfail · 1 deferred	✅ 14 pass · 119 xfail
streamgapdf	✅ 28 pass · 2 xfail	✅ 27 pass · 3 xfail
diskdf	✅ 129 pass · 0 xfail	✅ 112 pass · 17 xfail
streamdf + streamspraydf + streamTrack	✅ 142 pass · 17 xfail	✅ 66 pass · 93 xfail

Per-shard counts

Test shard	backend	pass	xfail	deferred	XPASS	fail	error
actionAngle	jax	112	89	0	0	0	0
actionAngle	torch	26	175	0	0	0	0
sphericaldf	jax	164	26	28	0	0	0
sphericaldf	torch	8	210	0	0	0	0
conversion + util + misc	jax	85	6	1	0	0	0
conversion + util + misc	torch	41	51	0	0	0	0
potential + scf + multipole	jax	—	—	—	—	—	—
potential + scf + multipole	torch	—	—	—	—	—	—
quantity + coords	jax	287	49	0	0	0	0
quantity + coords	torch	189	147	0	0	0	0
orbit (energy/Jacobi + from_name)	jax	0	0	115	0	0	0
orbit (energy/Jacobi + from_name)	torch	63	52	0	0	0	0
orbit + orbits (main)	jax	0	0	578	0	0	0
orbit + orbits (main)	torch	248	327	0	0	0	0
evolveddiskdf	jax	35	0	0	0	0	0
evolveddiskdf	torch	32	3	0	0	0	0
jeans + dynamfric	jax	17	2	2	0	0	0
jeans + dynamfric	torch	7	13	1	0	0	0
qdf + pv2qdf + streamgapdf_impulse + noninertial	jax	57	75	1	0	0	0
qdf + pv2qdf + streamgapdf_impulse + noninertial	torch	14	119	0	0	0	0
streamgapdf	jax	28	2	0	0	0	0
streamgapdf	torch	27	3	0	0	0	0
diskdf	jax	129	0	0	0	0	0
diskdf	torch	112	17	0	0	0	0
streamdf + streamspraydf + streamTrack	jax	142	17	0	0	0	0
streamdf + streamspraydf + streamTrack	torch	66	93	0	0	0	0