docs: record the Aris large-mask before/after witness tables

campaigns.yaml

@@ -45,7 +45,7 @@ campaigns:
   title: Interferometer likelihood and streaming
   status: active
   updated: '2026-10-15'
-  next: 'Three PyAutoArray tasks filed 2026-10-15 from the Aris 512×512 profiling (PyAutoPulse#62): #626 sparse data vector memory (active, --auto), #627 AdaptSplit log-det backend parity and #628 numba route-gate cost model (queued behind the PyAutoArray claim). Streaming 1e8 CPU/A100 rows and the #332 memo after-measurement still need authorization; curvature preload needs the fixed-mapper science decision; FFT levers remain unstarted.'
+  next: 'All three PyAutoArray members from the Aris 512×512 profiling (PyAutoPulse#62) merged 2026-10-15, pending release: #630 sparse data vector memory, #631 AdaptSplit log-det backend parity, #632 numba route-gate cost model; phase-2 autolens_profiling record in flight. Streaming 1e8 CPU/A100 rows and the #332 memo after-measurement still need authorization; curvature preload needs the fixed-mapper science decision; FFT levers remain unstarted.'
   evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/4db4d988e520a82219152743748c3188d485a4ba/wiki/campaigns/interferometer_likelihood.md
 - id: linear-solver
   title: Linear-solver accuracy and cost
@@ -307,7 +307,7 @@ tasks:
   priority: high
   path: tasks/interferometer_large_mask_aris.md
   updated: '2026-10-15'
-  next: 'PyAutoArray#626 active under --auto (worktree interferometer-sparse-data-vector; witness 7.2 GB -> <= ~3.5 GB at 5,746 vertices); #627 and #628 queued in Mind planned.md behind the PyAutoArray claim, in that order. Record each before/after witness table in the task file at close-out.'
+  next: 'All three library members merged (PRs #630, #631, #632; pending release). In flight: autolens_profiling phase-2 record (routing table under the cost model + Aris log-det after-measurement). Then copy that table here and close the task.'
   issue: https://github.com/PyAutoLabs/PyAutoPulse/issues/62
   evidence: https://github.com/PyAutoLabs/PyAutoArray/issues/626
 - id: interferometer_w_tilde_fft_size_levers

dashboard.md

@@ -1,6 +1,6 @@
 # PyAutoPulse — profiling dashboard
 
-<!-- pulse-campaigns:a53f6bbfc21f511541a477cdac96aaf7b677a7e058726ca4f839a61947a26133 -->
+<!-- pulse-campaigns:0ec1ffed93e62a62d70397b0e88007eb2e1d8284d3e4f54f10d7c51896a2ac7e -->
 ## Profiling Check In
 
 <details><summary>Full check-in prompt</summary>
@@ -94,7 +94,7 @@ Reviewed 2026-10-13. No campaign wiki/results change or new matching implementat
 
 <details><summary>Interferometer likelihood and streaming — active</summary>
 
-Reviewed 2026-10-15. Three PyAutoArray tasks filed 2026-10-15 from the Aris 512×512 profiling (PyAutoPulse#62): #626 sparse data vector memory (active, --auto), #627 AdaptSplit log-det backend parity and #628 numba route-gate cost model (queued behind the PyAutoArray claim). Streaming 1e8 CPU/A100 rows and the #332 memo after-measurement still need authorization; curvature preload needs the fixed-mapper science decision; FFT levers remain unstarted.
+Reviewed 2026-10-15. All three PyAutoArray members from the Aris 512×512 profiling (PyAutoPulse#62) merged 2026-10-15, pending release: #630 sparse data vector memory, #631 AdaptSplit log-det backend parity, #632 numba route-gate cost model; phase-2 autolens_profiling record in flight. Streaming 1e8 CPU/A100 rows and the #332 memo after-measurement still need authorization; curvature preload needs the fixed-mapper science decision; FFT levers remain unstarted.
 
 [Campaign evidence](https://github.com/PyAutoLabs/autolens_profiling/blob/4db4d988e520a82219152743748c3188d485a4ba/wiki/campaigns/interferometer_likelihood.md)
 
@@ -103,7 +103,7 @@ Reviewed 2026-10-15. Three PyAutoArray tasks filed 2026-10-15 from the Aris 512
 - [Interferometer fixed-mapper searches: reuse the W~ curvature matrix across likelihood calls via `preloads.curvature_matrix`](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_fixed_mapper_curvature_preload.md) — ready: Phase 1 Mind draft (2026-10-04) still unstarted; run it through start_dev. Human science call first - no production SLaM stage holds the mapper fixed today.
 - [fnnls warm-start memo: stop it slowing scattered evaluation streams (interferometer CPU, autolens_profiling#332)](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_nnls_memo_scattered_stream_guard.md) — active: Library half shipped: PyAutoArray#615 (closes #613) merged 2026-10-06, in 2026.10.7.1. Local solver-only witness (n=576): iid memo on/off 1.49x → 1.17x; walk 0.18x unchanged. Owed: re-run the autolens_profiling#332 ALMA Delaunay 2.17x after-measurement (needs compute authorization).
 - [Campaign: interferometer streaming (array-free) vs in-memory — memory and time scaling to 2e8 visibilities](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_streaming_scaling.md) — active: Phase 1 CPU cells + results/streaming_scaling rows merged (autolens_profiling#375, 2026-10-04); wiki/campaigns/interferometer_streaming.md. Optional: 1e8 CPU row and A100 rows (compute authorization). Library candidate unfiled: chunk transformer.image_from in apply_sparse_operator (memory).
-- [Interferometer large-mask likelihood (Aris 512×512): sparse data-vector memory, AdaptSplit log-det backend parity, numba route-gate cost model](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_large_mask_aris.md) — active: PyAutoArray#626 active under --auto (worktree interferometer-sparse-data-vector; witness 7.2 GB -&gt; &lt;= ~3.5 GB at 5,746 vertices); #627 and #628 queued in Mind planned.md behind the PyAutoArray claim, in that order. Record each before/after witness table in the task file at close-out.
+- [Interferometer large-mask likelihood (Aris 512×512): sparse data-vector memory, AdaptSplit log-det backend parity, numba route-gate cost model](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_large_mask_aris.md) — active: All three library members merged (PRs #630, #631, #632; pending release). In flight: autolens_profiling phase-2 record (routing table under the cost model + Aris log-det after-measurement). Then copy that table here and close the task.
 - [Interferometer W~ curvature matrix is FFT-bound on the mask extent: pruned padded FFT and real-space pixel scale on the A100](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_w_tilde_fft_size_levers.md) — ready: Unstarted (no PR/issue/branch/results, 2026-10-04). Select one bounded step when prioritised.
 
 </details>

tasks/interferometer_large_mask_aris.md

@@ -3,7 +3,7 @@
 Campaign: interferometer
 Task: interferometer_large_mask_aris
 Reviewed: 2026-10-15
-Status: active
+Status: active (all three library members merged 2026-10-15, pending release; phase-2 record and Aris log-det after-measurement in flight)
 
 Filed 2026-10-15 from the Aris interferometer profiling session (Delaunay + `AdaptSplit`,
 `use_jax=False` numba route and JAX jit, real-space masks 125×125 at 5,496 masked pixels and
@@ -42,3 +42,51 @@ scratchpad (temporary); the numbers above are the durable record.
 
 - Filed: #626 active (worktree `interferometer-sparse-data-vector`), #627 and #628 queued in Mind planned.md. Implementation worker running on #626; read-only root-cause investigation running for #627.
 - Next: after #626 PR-open and human `/prm`, start #627 (`/start_library`), then #628. Record each task's before/after witness table here on close-out.
+
+## Results 2026-10-15
+
+All three members merged on 2026-10-15 and are pending release. The tables below are copied
+from the PR bodies and the Mind completion records. Timings are not claimed: the machine's
+load average was 2–14 during measurement.
+
+### 1. Sparse data vector — [PyAutoArray#630](https://github.com/PyAutoLabs/PyAutoArray/pull/630) (merged 7cc71238)
+
+| run (Aris 512×512) | before | after |
+|---|---|---|
+| numba, 5,746 vertices, peak RSS | 7.17 GB | **3.14 GB** |
+| JAX jit, 1,459 vertices, peak RSS | 3.65 GB | 2.60 GB |
+| JAX, 5,746 vertices, XLA compile temp | 8.95 GB | 1.86 GB |
+| JAX jit, 5,746 vertices | infeasible | runs, peak RSS 3.24 GB |
+
+- The data vector matches the dense `mapping_matrix.T @ d~` to 5.8e-16 relative. The log likelihood moves by at most 2.5e-11 relative (≈3.5e-4 nats at |logL| ≈ 1.4e7).
+- Decision taken: the witness "logL unchanged to 1e-8" was read as fp64 precision, because an absolute 1e-8 is below float64 resolution at this logL. Revert: `git revert 8676b6a0`.
+
+### 2. AdaptSplit log-det backend parity — [PyAutoArray#631](https://github.com/PyAutoLabs/PyAutoArray/pull/631) (merged 0377eeaa)
+
+| check | before | after |
+|---|---|---|
+| synthetic ill-conditioned AdaptSplit Delaunay H (cond ≈ 2.2e9), NumPy vs JAX | different H | H parity 3.6e-15 |
+| same, log-det NumPy vs JAX | — | Δ 2.1e-9; both within 1.7e-9 of a 50-digit mpmath reference |
+| Aris 125×125 dump, log-det (investigation's simulated fix) | 2443.22 numba vs 2561.98 JAX | 2468.446318 on both |
+
+- Root cause: the log-det code was correct. `adaptive_pixel_signals_from` counted zero-weight mappings for data points lying on a vertex or edge, and scipy and the JAX Delaunay walk pick different triangles for those points. The two backends therefore built different H matrices.
+- Fix: zero-weight mappings are excluded for Delaunay and DelaunayNN only. The human chose this scope after review, because KNNBarycentric clips weights to zero over whole regions. KNN, KNNBarycentric and rectangular meshes are bit-identical.
+- Owed: the Aris setup was not re-run on the branch. The after-measurement goes into the phase-2 record (in flight as a separate autolens_profiling task).
+
+### 3. Numba route-gate cost model — [PyAutoArray#632](https://github.com/PyAutoLabs/PyAutoArray/pull/632) (merged 17caf228)
+
+| mask (extent) | vertices | nnz/col | measured faster | chosen (cost model) | old gate 60 |
+|---|---|---|---|---|---|
+| 512×512 (341²) | 1,459 | 187.6 | FFT | FFT ✓ | FFT |
+| 512×512 (341²) | 2,577 | 106.2 | numba | numba ✓ | FFT ✗ |
+| 512×512 (341²) | 5,746 | 47.7 | numba | numba ✓ | numba |
+| 125×125 (83²) | 1,405 / 2,947 / 5,526 | 11.7 / 5.6 / 3.0 | numba | numba ✓ | numba |
+
+- Cost model: `t_fft = c_fft · S · FFT work` (pocketfft `cost_guess` over the padded extent) and `t_dc = c_dc · nnz · ny · nx`, with c_fft 7.62e-10 and c_dc 1.77e-09. It was fitted on 9 paired timings and picks the faster route at 9/9, with residual ratios 0.68–1.47. Mask extents are 341² and 83²; 682 was the padded FFT grid.
+- Settings: `interferometer_numba_nnz_per_source_max` null means the cost model (new default), a positive value is a hard gate, and 0 turns numba off. The new `interferometer_route` setting forces `"numba"` or `"fft"`. Revert: set the gate back to 60.
+- The routing above is a live routing-only check. The timings come from the baseline table and were not re-run.
+
+### Next
+
+The autolens_profiling phase-2 record is in flight. It re-measures the routing table under the cost
+model and records the Aris log-det after-measurement. Copy that table here, then close the task.