Update source_plane regression to Richardson-converged truth

jax_profiling/point_source/results/image_plane_summary_v2026.4.13.6.json

@@ -7,14 +7,14 @@
     "positions_noise_sigma": 0.2,
     "free_parameters": 8
   },
-  "eager_per_call": 1.4605535200000304,
+  "eager_per_call": 2.956794569999329,
   "eager_log_likelihood": 0.3936326580483207,
-  "full_pipeline_single_jit": 0.029288589999850956,
+  "full_pipeline_single_jit": 0.04489453000060166,
   "full_pipeline_log_likelihood": 0.39363265804832137,
   "vmap": {
     "batch_size": 3,
-    "batch_time": 0.041269269999975225,
-    "per_call": 0.013756423333325074,
-    "speedup_vs_single_jit": 2.1
+    "batch_time": 0.07011199000044144,
+    "per_call": 0.02337066333348048,
+    "speedup_vs_single_jit": 1.9
   }
 }

jax_profiling/point_source/results/source_plane_summary_v2026.4.13.6.json

@@ -7,19 +7,19 @@
     "positions_noise_sigma": 0.2,
     "free_parameters": 8
   },
-  "eager_per_call": 0.0028598109999802544,
-  "eager_log_likelihood": -4496.798984131583,
+  "eager_per_call": 0.007736787000030745,
+  "eager_log_likelihood": -4491.83220547254,
   "full_pipeline_jits": true,
   "full_pipeline_blocker": null,
-  "full_pipeline_single_jit": 0.00038716000017302576,
+  "full_pipeline_single_jit": 0.0005248499997833278,
   "jit_able_prefix": {
     "name": "ray-trace observed positions to source plane",
-    "per_call": 0.00034986999999091494
+    "per_call": 0.0004281299996364396
   },
   "vmap_prefix": {
     "batch_size": 3,
-    "batch_time": 0.0002824099999997998,
-    "per_call": 9.413666666659993e-05,
-    "speedup_vs_single_jit_prefix": 3.7
+    "batch_time": 0.0004249800003890414,
+    "per_call": 0.00014166000012968047,
+    "speedup_vs_single_jit_prefix": 3.0
   }
 }

jax_profiling/point_source/source_plane.py

@@ -11,23 +11,6 @@
 ray-traced positions and the model source position.  No image-plane solver
 is required.
 
-LIBRARY-LEVEL NUMERICAL DRIFT (np vs jnp magnifications)
----------------------------------------------------------
-
-The earlier ``TracerArrayConversionError`` JIT blocker (``Grid2DIrregular.
-grid_2d_via_deflection_grid_from`` not propagating ``xp``) has been fixed
-upstream — the full source-plane likelihood now JIT-traces end-to-end.
-
-A separate numerical issue remains: the eager-np and eager-jnp paths of
-``LensCalc.magnification_2d_via_hessian_from`` disagree at ~5e-4 relative,
-which amplifies into a ~0.1% drift in the final log-likelihood.  See the
-follow-up prompt ``admin_jammy/prompt/autolens/lens_calc_magnification_xp_divergence.md``
-for the full analysis and plan.
-
-Until that lands, the JIT regression assertion below is loosened from
-``rtol=1e-4`` to ``rtol=2e-3`` so the script stays green.  The eager-only
-assertion remains tight.
-
 Pytree-native parameter inputs
 ------------------------------
 
@@ -496,14 +479,15 @@ def ray_trace_to_source_plane(params_tree, positions_raw):
 
 
 # ===================================================================
-# Regression assertions (eager-only — see BLOCKER above)
+# Regression assertions (eager and full-pipeline JIT)
 # ===================================================================
 #
 # Seeded simulator (noise_seed=1 in simulators/point_source.py) + prior-median
-# parameter vector make the eager source-plane log-likelihood deterministic.
-# The full-JIT regression assertion is gated on full_pipeline_jits — once the
-# library blocker is fixed, both branches will fire.
-EXPECTED_LOG_LIKELIHOOD_SOURCE_PLANE = -4496.798984131583
+# parameter vector make the source-plane log-likelihood deterministic. Both
+# the eager numpy and the full-pipeline JIT paths now agree to float64
+# precision, following the Richardson-extrapolation fix to
+# LensCalc.hessian_from in PyAutoGalaxy (PR #358).
+EXPECTED_LOG_LIKELIHOOD_SOURCE_PLANE = -4491.83220547254
 
 np.testing.assert_allclose(
     log_likelihood_ref,
@@ -520,13 +504,10 @@ def ray_trace_to_source_plane(params_tree, positions_raw):
 )
 
 if full_pipeline_jits:
-    # rtol loosened pending the np/jnp parity fix in
-    # LensCalc.magnification_2d_via_hessian_from — see
-    # admin_jammy/prompt/autolens/lens_calc_magnification_xp_divergence.md
     np.testing.assert_allclose(
         float(full_result),
         EXPECTED_LOG_LIKELIHOOD_SOURCE_PLANE,
-        rtol=2e-3,
+        rtol=1e-4,
         err_msg=(
             f"point_source/source_plane: regression — JIT log_likelihood drifted "
             f"(got {float(full_result)}, expected {EXPECTED_LOG_LIKELIHOOD_SOURCE_PLANE})"