Jammy2211 · Oct 15, 2026 · Oct 15, 2026 · Oct 15, 2026
diff --git a/campaigns.yaml b/campaigns.yaml
@@ -45,7 +45,7 @@ campaigns:
   title: Interferometer likelihood and streaming
   status: active
   updated: '2026-10-15'
-  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.'
+  next: 'Aris large-mask work (PyAutoPulse#62): #630/#631/#632 merged 2026-10-15, pending release; knowledge section and two human decisions recorded; phase-2 record at PR-open (autolens_profiling#420); L1 Cholesky reuse (PyAutoArray#633) in development, L2 direct_conv symmetry filed and unstarted. 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: '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.'
+  next: 'Knowledge written 2026-10-15 (cost structure, routing model, levers, collaborator settings) and the two human decisions recorded. #630/#631/#632 merged, pending release. Phase-2 record at PR-open (autolens_profiling#420): re-pin its table on merge. L1 Cholesky reuse (PyAutoArray#633) in development; L2 direct_conv symmetry filed, unstarted.'
   issue: https://github.com/PyAutoLabs/PyAutoPulse/issues/62
   evidence: https://github.com/PyAutoLabs/PyAutoArray/issues/626
 - id: interferometer_w_tilde_fft_size_levers

diff --git a/dashboard.html b/dashboard.html
diff --git a/dashboard.md b/dashboard.md
@@ -1,6 +1,6 @@
 # PyAutoPulse — profiling dashboard
 
-<!-- pulse-campaigns:0ec1ffed93e62a62d70397b0e88007eb2e1d8284d3e4f54f10d7c51896a2ac7e -->
+<!-- pulse-campaigns:4e132b2d2378f34b9712646cb9a497b82be62a3d7d484f8136ab2bc19cc70303 -->
 ## 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. 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.
+Reviewed 2026-10-15. Aris large-mask work (PyAutoPulse#62): #630/#631/#632 merged 2026-10-15, pending release; knowledge section and two human decisions recorded; phase-2 record at PR-open (autolens_profiling#420); L1 Cholesky reuse (PyAutoArray#633) in development, L2 direct_conv symmetry filed and unstarted. 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. All three PyAutoArray members from the Aris 512×512 profil
 - [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: 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 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: Knowledge written 2026-10-15 (cost structure, routing model, levers, collaborator settings) and the two human decisions recorded. #630/#631/#632 merged, pending release. Phase-2 record at PR-open (autolens_profiling#420): re-pin its table on merge. L1 Cholesky reuse (PyAutoArray#633) in development; L2 direct_conv symmetry filed, unstarted.
 - [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>
@@ -330,7 +330,9 @@ Project `autofit_profiling`, scope `pyautofit-profiling`, read from [PyAutoLabs/
 </details>
 
 
-<!-- decision-history:8f1840916151d6a62b1cb1cd473edede0a3647c198fe8ef2371ce31a66c873a5 -->
+<!-- decision-history:20f0256fc1d56eda82a910f81ddb71d393e76eeaf26e486d9c836b00282b99f9 -->
 ## Decision History
 
+- <a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/interferometer-route-gate-cost-model.md">Route interferometer curvature by a calibrated two-constant cost model</a>
+- <a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/adaptsplit-zero-weight-delaunay-only.md">Exclude zero-weight mappings from AdaptSplit pixel signals on Delaunay meshes only</a>
 - <a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/linear-solver-mapper-default-jacobi.md">Keep Jacobi PDIP as the Mapper default and raw+polish as the MGE-only default</a>
diff --git a/decisions.yaml b/decisions.yaml
@@ -4,3 +4,11 @@ decisions:
   title: Keep Jacobi PDIP as the Mapper default and raw+polish as the MGE-only default
   date: '2026-10-08'
   url: https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/linear-solver-mapper-default-jacobi.md
+- id: adaptsplit-zero-weight-delaunay-only
+  title: Exclude zero-weight mappings from AdaptSplit pixel signals on Delaunay meshes only
+  date: '2026-10-15'
+  url: https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/adaptsplit-zero-weight-delaunay-only.md
+- id: interferometer-route-gate-cost-model
+  title: Route interferometer curvature by a calibrated two-constant cost model
+  date: '2026-10-15'
+  url: https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/interferometer-route-gate-cost-model.md
diff --git a/decisions/adaptsplit-zero-weight-delaunay-only.md b/decisions/adaptsplit-zero-weight-delaunay-only.md
@@ -0,0 +1,113 @@
+# Exclude zero-weight mappings from AdaptSplit pixel signals on Delaunay meshes only
+
+Date: 2026-10-15
+Decision maker: James Nightingale
+Decision source: AskUserQuestion answer "Delaunay only (Recommended by review)" in session 3e3dea38, 2026-10-15. Quoted on [PyAutoArray#627 (issuecomment-6200058449)](https://github.com/PyAutoLabs/PyAutoArray/issues/627#issuecomment-6200058449). Recorded on the human's explicit direction in the same session ("Yes, record it as my decision").
+
+## Campaign
+
+This decision belongs to the interferometer campaign, task `interferometer_large_mask_aris`
+([PyAutoPulse#62](https://github.com/PyAutoLabs/PyAutoPulse/issues/62)). The task profiles the Aris
+SPT2319-55 interferometer setup: a Delaunay mesh with `AdaptSplit` regularization, on real-space
+masks of 125×125 (5,496 px) and 512×512 (91,507 px), with 5.5M visibilities.
+
+The numba/NumPy and JAX backends returned different log evidences at identical parameters. The whole
+difference was in `log_det_regularization_matrix_term`: 2443.22 on numba against 2561.98 on JAX for
+one 125×125 vector. Evidences were therefore not comparable across backends or across pipeline phases
+that mix them. The issue was
+[PyAutoArray#627](https://github.com/PyAutoLabs/PyAutoArray/issues/627).
+
+## Evidence
+
+The evidence is at PyAutoArray `ccddfba6` (before the fix) and `0377eeaa` (the merge of
+[#631](https://github.com/PyAutoLabs/PyAutoArray/pull/631)), on the laptop in fp64. It is drawn from
+the #627 comments "Root cause found" (issuecomment-6197915508) and "Parked at the review leg"
+(issuecomment-6199913182), and from Mind `complete/2026/10/adaptsplit-logdet-backend-parity.md`.
+
+- **The factorizations were not the cause.**
+  - On each backend's own H, five methods agree with a 50-digit mpmath reference to ≤ 2e-8: NumPy
+    dense Cholesky, SuperLU, JAX Cholesky, slogdet and diagonal scaling.
+  - The two backends built different H matrices: max |H_np − H_jax| = 3.21, on max |H| 31.9.
+- **Root cause.**
+  - `adaptive_pixel_signals_from` divided each pixel's summed signal by a count that included
+    zero-weight mapping entries.
+  - The 125×125 mesh is a stride-2 subsample of the data grid, so 1,375 data points sit exactly on
+    vertices. For such a point any surrounding triangle is valid.
+  - scipy `find_simplex` and the JAX Delaunay walk chose different triangles in 1,114 of 5,496 rows.
+    Both choices are geometrically correct.
+  - The phantom zero-weight neighbours changed the counts, then the pixel signals (max diff 0.23),
+    the AdaptSplit weights (0.53) and H.
+  - Chi², the curvature matrix, the data vector and the curvature log det were unaffected.
+- **Simulated fix: count only non-zero weights.** Signals agree to 4e-15 and H to 4.8e-13. The log
+  det is 2468.446318 on both backends (Δ 4e-8).
+- **Review finding on the first patch, which applied the rule to every interpolator.**
+  - `KNNBarycentric` clips negative barycentric weights to exactly 0 for every query outside its
+    3-nearest triangle, so whole regions are affected, not only coincident points.
+  - Excluding those zeros would change its pixel signals and evidence broadly.
+  - It would also add a jump in the likelihood: moving one weight from 1e-9 to 0 moved a test scalar
+    from 6.2358 to 6.4889, and the gradient at w = 0 is exactly 0.
+  - Rectangular bilinear meshes are affected only in exact-coincidence cases.
+- **Shipped result (#631, Delaunay-only scope).**
+  - On a synthetic ill-conditioned AdaptSplit case (cond ≈ 2.2e9), H parity is 3.6e-15 and the
+    log dets are 2.1e-9 apart, both within 1.7e-9 of a 50-digit reference.
+  - The independent re-review found KNN, KNNBarycentric and rectangular bit-identical to main.
+- **After-measurement on the Aris data.**
+  - Source: [autolens_profiling#420](https://github.com/PyAutoLabs/autolens_profiling/pull/420), head
+    `32372865`, **at PR-open, not merged**, on PyAutoArray `17caf228`.
+  - numba and JAX agree to ≤ 4.4e-8 at all six cells (125: 1,405/2,947/5,526 vertices; 512:
+    1,459/2,577/5,746).
+  - The original 125×125 vector gives 2468.4463180 on both backends.
+- **Limitation: a residual remains.** The split regularization builders add an absolute 1e-8 diagonal
+  jitter. Once max(coeff)^4 ≳ 1e6, the log det is round-off-dominated (cond 1e16–1e17).
+  - In 20 prior draws on 125×125 with the counting fix, 6 still differed by 5e-3–2.2.
+  - This decision does not address the residual, and no work on it is filed.
+
+## Options considered
+
+- **(a) Accept the rule for every mesh, including KNN and KNNBarycentric.** This is the simplest
+  code. But KNNBarycentric evidence would change across whole regions, and its likelihood would gain
+  a discontinuity on a mesh used for gradient-based fits.
+- **(b) Delaunay only** (chosen; recommended by the independent review). Exclude zero-weight mappings
+  for Delaunay and DelaunayNN, and keep the old counting elsewhere. It fixes the parity bug where it
+  arises, at the cost of one keyword and one interpolator class attribute.
+- **(c) Weight-normalised signals.** Divide the summed signal by the summed weights rather than by a
+  count, so zero weights drop out on every mesh without a jump. This was raised in the session as
+  the route if KNN-family parity is ever needed. It was not prototyped or measured.
+  - The #627 park comment listed a different third option: make the scipy/JAX triangle choice
+    deterministic for coincident points and keep the count semantics. That option had been rejected
+    earlier in the investigation.
+
+## Decision
+
+Exclude zero-weight mapping entries from the adaptive pixel-signal counts for Delaunay and
+DelaunayNN interpolators only. KNN, KNNBarycentric and rectangular meshes keep the previous counting.
+
+## Implications
+
+- **Shipped in [#631](https://github.com/PyAutoLabs/PyAutoArray/pull/631)** (merge `0377eeaa`,
+  pending release).
+  - `adaptive_pixel_signals_from(..., exclude_zero_weight_mappings=False)` gains a keyword.
+  - It is driven by the interpolator class attribute `exclude_zero_weight_mappings`: True on
+    InterpolatorDelaunay and InterpolatorDelaunayNN, False on the base interpolator and KNN.
+- **Evidence values change, the same way on both backends,** on Delaunay and DelaunayNN meshes where
+  data points sit exactly on vertices or edges, for example Overlay or grid-subsample image meshes.
+  Any pinned evidence for such a mesh will shift. No smoke script failed (151/151), and no workspace
+  change was needed.
+- **KNN, KNNBarycentric and rectangular meshes are bit-identical** to the previous release.
+- numba/JAX evidences on Delaunay AdaptSplit fits are now comparable to ~1e-8 at moderate
+  coefficients. The high-coefficient jitter regime is still backend-sensitive.
+
+## Revisit conditions
+
+- **KNN-family parity is needed,** for example a KNNBarycentric backend-parity requirement. Then
+  evaluate weight-normalised signals (option c), which avoids the discontinuity of option (a).
+- **The jitter residual matters in practice.** That is, fits reach max(coeff)^4 ≳ 1e6, or the 512×512
+  non-PD rate (~1 in 4 draws before #631) is still seen. A relative jitter
+  (`1e-8 * max(regularization_weights**2)`) is the candidate. It changes evidence values, so it needs
+  its own decision.
+- **A change to the Delaunay point-location routines,** in scipy or the JAX walk, or a new
+  interpolator whose weights can be exactly zero.
+
+## History
+
+This is the first decision on this question. It supersedes nothing.