docs: cluster-pointsolver phase-1 knowledge + A2744 pilot pull check-in 2026-10-23

campaigns.yaml

@@ -1,5 +1,5 @@
 version: 1
-last_checkin: '2026-10-17T15:12:59Z'
+last_checkin: '2026-10-23T15:00:32Z'
 campaigns:
 - id: setup-baseline
   title: Setup baseline collection and scientific acceptance
@@ -47,13 +47,9 @@ campaigns:
 - id: cluster-pointsolver
   title: Cluster PointSolver robustness and performance
   status: active
-  updated: '2026-10-17'
-  next: 'No campaign wiki/results change or new matching implementation claim found. Representative cluster baseline
-    and observable-overflow policy remain unstarted. Choose one bounded phase; inherited single-source results
-    do not qualify a representative cluster baseline. Check-in 2026-10-17: no project wiki/results/issue change
-    since 2026-10-13 beyond autolens_profiling#420/#422/#424 (interferometer only); releases 2026.10.16.1/.2 and
-    2026.10.17.1 supply no new measurement.'
-  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/4db4d988e520a82219152743748c3188d485a4ba/wiki/campaigns/cluster_pointsolver.md
+  updated: '2026-10-23'
+  next: 'Human decision 2026-10-23 (chat): keep the 1.0″ starting grid as the production default (decision record cluster-pointsolver-starting-grid; closes the phase-1 deferral). Inputs: phase-1 simple sweep (#438), A2744 pilot job 412573 (1.0″ misses fewest, 2/69), starting-grid follow-up autolens_profiling#439 → #440 merged 03d36c81 (no fixed scale finds every image; no separation rule). Mitigation: the workspace validation check. Next bounded task: spurious-core candidates library task (start_dev in progress), then the validation check. No default changed.'
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/03d36c810b2a4f7d31342d9a75f46cd7fe10f43a/results/notes/cluster_pointsolver_campaign.md
 - id: interferometer
   title: Interferometer likelihood and streaming
   status: parked
@@ -285,12 +281,11 @@ tasks:
 - id: cluster_pointsolver_speed
   campaign: cluster-pointsolver
   title: Cluster PointSolver — robustness, analysis settings and performance
-  status: ready
+  status: active
   priority: normal
   path: tasks/cluster_pointsolver_speed.md
-  updated: '2026-10-13'
-  next: 'Unissued, no claim. Pick one bounded phase per Mind epics.md: observable-overflow invalid-result policy
-    + CI witness, or a representative-data baseline cell.'
+  updated: '2026-10-23'
+  next: 'Decision taken 2026-10-23 (human): keep 1.0″ + validation check; #439/#440 merged (03d36c81). Next: spurious-core candidates library task (start_dev in progress), then the workspace validation check.'
 - id: imaging_production_over_sampling
   campaign: imaging
   title: Profile imaging pixelizations at production over-sampling
@@ -384,9 +379,8 @@ tasks:
   status: ready
   priority: normal
   path: tasks/point_solver_profiling_cells.md
-  updated: '2026-10-13'
-  next: 'Unissued, no claim. Candidate first cell: a representative-data cluster likelihood_runtime baseline
-    (single-source preset).'
+  updated: '2026-10-23'
+  next: 'Unissued, no claim. Candidate first cell: a representative-data cluster likelihood_runtime baseline (single-source preset). Reconcile onto the phase-1 harness (scripts/cluster/image_plane/robust_settings_sweep.py) when the next cell is issued; no duplicate cell.'
 - id: point_source_image_plane_gpu_breakdown
   campaign: point-source-image-gpu
   title: 'Point-source A100 speed-up campaign: profile and optimize with the shared breakdown'

checkins/2026-10-23.md

@@ -0,0 +1,33 @@
+# Profiling check-in — 2026-10-23 (directed: cluster-pointsolver)
+
+Reviewed at 2026-10-23T13:54:43Z. This check-in covered only the **cluster-pointsolver** campaign at the
+human's direction; other campaigns were not reviewed and their ledger rows are unchanged.
+Review time and board generation are not measurement times. It authorizes no compute, default
+change, resubmission, scientific acceptance, release or merge.
+
+## Verified availability
+
+- RAL was reachable (SSH `euclid_jump`). `sacct` for job 412573 and its log (1951 lines) were
+  read. No compute was run and no results were collected into the project.
+- GitHub: autolens_profiling#438 merged at `c8c9abb3` (2026-10-22); Mind record
+  `complete/2026/10/cluster-pointsolver-robust-settings.md`.
+
+## Campaign review
+
+| Campaign | Changes since 2026-10-17 | Current state | Blockers | Next bounded step |
+|---|---|---|---|---|
+| [Cluster PointSolver robustness and performance](../tasks/cluster_pointsolver_speed.md) | Phase 1 merged ([#438](https://github.com/PyAutoLabs/autolens_profiling/pull/438), `c8c9abb3`); human decision DEFERRED 2026-10-22 (default stays 1.0″); Knowledge block landed. A2744 pilot (job 412573) pulled: **FAILED** after 16 diagnostic rows (03:33:07, MaxRSS 85.5 GB) on LLVM `Cannot allocate memory`; no JSON written. A2744 coverage: 1.0″ 67/69, 0.5″ 66/69, 0.7″ 64/69, 2.0″ 61/69, truth-bbox 58–60/69 — opposite ordering to `simple`, so the scale ranking is dataset-dependent. RSS 8.3 → 39.4 GB with live arrays ~0: retained compiled executables plus compile transients. | active | The A2744 sweep harness must release executables between rows (or run one row per process) before any A2744 full sweep; resubmission needs explicit human authorization (none given). Phase 2 CI witnesses: nothing accepted to distil. | Starting-grid / close-member-image Mind draft (`draft/research/autolens_profiling/cluster_pointsolver_starting_grid_close_member_image.md`) via start_dev; one task, no bulk queue. |
+
+## Task changes
+
+| Task | State | Note |
+|---|---|---|
+| [cluster_pointsolver_speed](../tasks/cluster_pointsolver_speed.md) | ready → active | Check-in 2026-10-23 section, phase-1 Knowledge block and A2744 pilot paragraph added. |
+| [point_solver_profiling_cells](../tasks/point_solver_profiling_cells.md) | ready | Reconciles onto phase 1's harness (`scripts/cluster/image_plane/robust_settings_sweep.py`) when the next cell is issued; no duplicate cell. |
+
+## Unavailable or unverified
+
+- The A2744 model is a synthetic prior-median model with 3 judged draws; its rows are diagnostic
+  only and carry no speed claim. The backend rows and timing rounds were never reached.
+- The pilot log is not yet tracked in autolens_profiling; landing it is deliverable 0 of the
+  starting-grid draft.

dashboard.md

@@ -1,6 +1,6 @@
 # PyAutoPulse — profiling dashboard
 
-<!-- pulse-campaigns:4e7b5e1126147b4593a7e870235fd9a5d367256202aaeb9408c0254298791675 -->
+<!-- pulse-campaigns:b2279cce088381119a25ae69e4a1a493c347db477e562fe28b38596ca10f6197 -->
 ## Profiling Check In
 
 <details><summary>Full check-in prompt</summary>
@@ -25,7 +25,7 @@ After taking action, report what changed, what the evidence supports and what re
 
 </details>
 
-Last check-in: 2026-10-17T15:12:59Z (review date, not measurement freshness).
+Last check-in: 2026-10-23T15:00:32Z (review date, not measurement freshness).
 
 ## Active campaigns
 
@@ -68,14 +68,14 @@ Reviewed 2026-10-17. Parked as research, human decision 2026-10-17 (chat) (&quot
 
 <details><summary>Cluster PointSolver robustness and performance — active</summary>
 
-Reviewed 2026-10-17. No campaign wiki/results change or new matching implementation claim found. Representative cluster baseline and observable-overflow policy remain unstarted. Choose one bounded phase; inherited single-source results do not qualify a representative cluster baseline. Check-in 2026-10-17: no project wiki/results/issue change since 2026-10-13 beyond autolens_profiling#420/#422/#424 (interferometer only); releases 2026.10.16.1/.2 and 2026.10.17.1 supply no new measurement.
+Reviewed 2026-10-23. Human decision 2026-10-23 (chat): keep the 1.0″ starting grid as the production default (decision record cluster-pointsolver-starting-grid; closes the phase-1 deferral). Inputs: phase-1 simple sweep (#438), A2744 pilot job 412573 (1.0″ misses fewest, 2/69), starting-grid follow-up autolens_profiling#439 → #440 merged 03d36c81 (no fixed scale finds every image; no separation rule). Mitigation: the workspace validation check. Next bounded task: spurious-core candidates library task (start_dev in progress), then the validation check. No default changed.
 
-[Campaign evidence](https://github.com/PyAutoLabs/autolens_profiling/blob/4db4d988e520a82219152743748c3188d485a4ba/wiki/campaigns/cluster_pointsolver.md)
+[Campaign evidence](https://github.com/PyAutoLabs/autolens_profiling/blob/03d36c810b2a4f7d31342d9a75f46cd7fe10f43a/results/notes/cluster_pointsolver_campaign.md)
 
 ### Active tasks
 
-- [Cluster PointSolver — robustness, analysis settings and performance](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/cluster_pointsolver_speed.md) — ready: Unissued, no claim. Pick one bounded phase per Mind epics.md: observable-overflow invalid-result policy + CI witness, or a representative-data baseline cell.
-- [PointSolver profiling cells: lensed quasar → cluster runtime tier → single/multi-source → multiplane](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/point_solver_profiling_cells.md) — ready: Unissued, no claim. Candidate first cell: a representative-data cluster likelihood_runtime baseline (single-source preset).
+- [Cluster PointSolver — robustness, analysis settings and performance](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/cluster_pointsolver_speed.md) — active: Decision taken 2026-10-23 (human): keep 1.0″ + validation check; #439/#440 merged (03d36c81). Next: spurious-core candidates library task (start_dev in progress), then the workspace validation check.
+- [PointSolver profiling cells: lensed quasar → cluster runtime tier → single/multi-source → multiplane](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/point_solver_profiling_cells.md) — ready: Unissued, no claim. Candidate first cell: a representative-data cluster likelihood_runtime baseline (single-source preset). Reconcile onto the phase-1 harness (scripts/cluster/image_plane/robust_settings_sweep.py) when the next cell is issued; no duplicate cell.
 
 </details>
 
@@ -420,9 +420,10 @@ Project `autolens_inference`, scope `None`, read from [PyAutoLabs/autolens_infer
 </details>
 
 
-<!-- decision-history:731ca0edc0c01337b75841f88d79bd1838b31ba05910928a6353dc43150ca0ea -->
+<!-- decision-history:4d88bb773a36f774562f9e06d52a024b839cc44d2aa91282311891f6064c43c3 -->
 ## Decision History
 
+- <a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/cluster-pointsolver-starting-grid.md">Keep the 1.0″ PointSolver starting grid for cluster fits, with a validation check</a>
 - <a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/direct-conv-symmetric-gate.md">Gate the symmetric direct_conv kernel at work fraction 0.7, with ymin column ordering</a>
 - <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>

decisions.yaml

@@ -16,3 +16,7 @@ decisions:
   title: Gate the symmetric direct_conv kernel at work fraction 0.7, with ymin column ordering
   date: '2026-10-16'
   url: https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/direct-conv-symmetric-gate.md
+- id: cluster-pointsolver-starting-grid
+  title: Keep the 1.0″ PointSolver starting grid for cluster fits, with a validation check
+  date: '2026-10-23'
+  url: https://github.com/PyAutoLabs/PyAutoPulse/blob/main/decisions/cluster-pointsolver-starting-grid.md

decisions/cluster-pointsolver-starting-grid.md

@@ -0,0 +1,80 @@
+# Keep the 1.0″ PointSolver starting grid for cluster fits, with a validation check
+
+Date: 2026-10-23
+Decision maker: James Nightingale
+Decision source: Human decision in chat on 2026-10-23 ("keep the 1.0″ starting grid"), taken after the phase-1 decision was DEFERRED on 2026-10-22 pending the starting-grid follow-up and the A2744 pull. Recorded in Mind [complete/2026/10/cluster-pointsolver-starting-grid.md](https://github.com/PyAutoLabs/PyAutoMind/blob/main/complete/2026/10/cluster-pointsolver-starting-grid.md).
+
+## Campaign
+
+This decision belongs to the cluster PointSolver campaign, task
+[`cluster_pointsolver_speed`](../tasks/cluster_pointsolver_speed.md). Phase 1
+([autolens_profiling#436](https://github.com/PyAutoLabs/autolens_profiling/issues/436) →
+[#438](https://github.com/PyAutoLabs/autolens_profiling/pull/438)) measured which triangle-solver
+settings find every multiple image of cluster-scale point sources. The open question was the
+step-0 initial grid scale: the production 1.0″ against a finer start.
+
+Every result below is judged against an independent oracle that samples the multi-plane
+lens-equation residual densely and Newton-polishes each root (calibrated to 2.84e-8″). It is not
+the triangle solver.
+
+## Evidence
+
+Ledger: [results/notes/cluster_pointsolver_campaign.md at `03d36c81`](https://github.com/PyAutoLabs/autolens_profiling/blob/03d36c810b2a4f7d31342d9a75f46cd7fe10f43a/results/notes/cluster_pointsolver_campaign.md).
+The three inputs use different models and hardware; their numbers are not interchangeable.
+
+- **`simple` sweep (phase 1).** Synthetic `dataset/cluster/simple`: 13 components, 3 planes,
+  2 sources, 101 judged draws. RAL job 412572, Xeon 8490H, CPU fp64, production extent
+  100×100 @ 1.0″. All rows cover 306/306 prior roots. Of 300 stress roots, 0.5″ and 0.7″ miss 2,
+  the production 1.0″ misses 5 and 2.0″ misses 6. Warmed two-source solve: 1.0″ 5.29 ms, 0.5″
+  6.82 ms (0.78× speed), 0.7″ 7.02 ms (0.75×). No row is admissible. Mind record
+  [cluster-pointsolver-robust-settings.md](https://github.com/PyAutoLabs/PyAutoMind/blob/main/complete/2026/10/cluster-pointsolver-robust-settings.md).
+- **A2744 pilot.** A2744-geometry synthetic model at prior medians (not a fitted cluster model),
+  7 sources, 3 judged draws, 69 oracle roots. RAL job 412573, CPU fp64, pulled 2026-10-23. The job
+  FAILED after 16 diagnostic rows on LLVM memory exhaustion, so there is no timing claim. Coverage:
+  1.0″ misses 2/69, 0.5″ 3, 0.7″ 5, 2.0″ 8. The ordering is the opposite of `simple`.
+- **Starting-grid follow-up** ([#439](https://github.com/PyAutoLabs/autolens_profiling/issues/439) →
+  [#440](https://github.com/PyAutoLabs/autolens_profiling/pull/440), merged `03d36c81`; Mind
+  [cluster-pointsolver-starting-grid.md](https://github.com/PyAutoLabs/PyAutoMind/blob/main/complete/2026/10/cluster-pointsolver-starting-grid.md)).
+  autolens_visualization cluster model, laptop CPU fp64, 17 judged draws, diagnostic timings.
+  The workspace 1.0″ grid loses `point_0`'s central image 0.39″ from `main_1`'s core, and so does
+  every scale down to 0.3″. Only 0.2″ finds it (60.5 ms vs 7.4 ms warmed at 1.0″, ~8×), and 0.2″
+  still misses 3/108 roots. Six draw-sources are found at a coarser scale and lost at a finer one.
+  No single-separation rule holds over 236 draw-sources. The solver also returns spurious
+  candidates (residual 0.04″) on `r_core = 0` member cores.
+
+## Options considered
+
+- **0.5″ start.** Fewest misses on `simple` (2) at ~0.77× speed, but 3 misses on A2744 (more than
+  1.0″) and it does not find the viz model's central image.
+- **Fixed 0.2″ start.** ~8× the per-solve cost (laptop) and still misses 3/108 viz roots.
+- **Separation-based scale rule.** Unsupported: "scale ≲ separation / 1.06" is contradicted, and
+  the only uncontradicted form rests on two synthetic models.
+- **Keep 1.0″ and add a validation check.** Every observed image must have a solved counterpart;
+  otherwise re-solve on a small fine grid around it and say so. Drop candidates whose
+  source-plane residual exceeds the solver tolerance.
+
+## Decision
+
+Keep the 1.0″ starting grid as the production default, and adopt the validation check as the
+mitigation for missed images. This closes the phase-1 decision deferred on 2026-10-22.
+
+## Implications
+
+No library or workspace default is changed by this record. The validation check is a Mind draft
+(`draft/feature/autolens_workspace/cluster_point_solver_validation_check.md`, unissued). The
+spurious-core candidates are a separate library task
+(`draft/feature/autolens/point_solver_spurious_core_candidates.md`, start_dev in progress), next.
+The solver is not in the likelihood (`FitPositionsSourceSolved` traces the observed images), so a
+costly check at validation/visualization time does not slow fits.
+
+## Revisit conditions
+
+Evidence comes from synthetic models only; there is no fitted A2744 reference tracer, and the
+A2744 pilot covers 3 draws without timings. A full A2744 sweep (needs a reference tracer and
+human authorization), a fitted cluster model, or a validation check that fires often in practice
+would reopen the choice.
+
+## History
+
+First record for this choice. It follows the phase-1 deferral (2026-10-22) recorded in the task
+and the profiling ledger.