docs: prepare HST MGE pilot collection choices

campaigns.yaml

@@ -5,10 +5,9 @@ campaigns:
   title: Setup baseline collection and scientific acceptance
   status: needs-decision
   updated: '2026-10-13'
-  next: 'HST/MGE raw fixed-parameter CPU runner merged in #416; Pulse handoff #55 merged and Mind implementation
-    closed. Resolve exact production host/stack/data, full settings, independent correctness witness/tolerances
-    and background-load method. No collection, campaign freeze or acceptance authorized; global compile dispatch
-    remains unavailable.'
+  next: 'Review the Pulse collection choices (candidate input hashes and 31-worker resource plan verified without
+    execution) alongside project correctness audit #417. Resolve data origin/suitability, named host/stack/full
+    settings, witness/tolerances and background-load method. No freeze, compute or acceptance authorized.'
   evidence: https://github.com/PyAutoLabs/autolens_profiling/pull/416
 - id: point-source-image-cpu
   title: Point-source image plane · CPU
@@ -122,10 +121,9 @@ tasks:
   priority: high
   path: tasks/hst_mge_cpu_pilot.md
   updated: '2026-10-13'
-  next: 'HST/MGE raw fixed-parameter CPU runner merged in #416; Pulse handoff #55 merged and Mind implementation
-    closed. Resolve exact production host/stack/data, full settings, independent correctness witness/tolerances
-    and background-load method. No collection, campaign freeze or acceptance authorized; global compile dispatch
-    remains unavailable.'
+  next: 'Review the Pulse collection choices (candidate input hashes and 31-worker resource plan verified without
+    execution) alongside project correctness audit #417. Resolve data origin/suitability, named host/stack/full
+    settings, witness/tolerances and background-load method. No freeze, compute or acceptance authorized.'
   issue: https://github.com/PyAutoLabs/autolens_profiling/issues/415
   evidence: https://github.com/PyAutoLabs/autolens_profiling/pull/416
 - id: runtime_cell_single_jit_gpu_warmup

dashboard.md

@@ -1,6 +1,6 @@
 # PyAutoPulse — profiling dashboard
 
-<!-- pulse-campaigns:fed2012227f4b2e256c8788553e72329af859ee4b89aaf71bf85337ed475948c -->
+<!-- pulse-campaigns:d39fa8eaef6911cdb8a186ab8544a71e3ace4da6f8ef385069edf3cba0c43878 -->
 ## Profiling Check In
 
 <details><summary>Full check-in prompt</summary>
@@ -31,14 +31,14 @@ Last check-in: 2026-10-13T17:51:40Z (review date, not measurement freshness).
 
 <details><summary>Setup baseline collection and scientific acceptance — needs-decision</summary>
 
-Reviewed 2026-10-13. HST/MGE raw fixed-parameter CPU runner merged in #416; Pulse handoff #55 merged and Mind implementation closed. Resolve exact production host/stack/data, full settings, independent correctness witness/tolerances and background-load method. No collection, campaign freeze or acceptance authorized; global compile dispatch remains unavailable.
+Reviewed 2026-10-13. Review the Pulse collection choices (candidate input hashes and 31-worker resource plan verified without execution) alongside project correctness audit #417. Resolve data origin/suitability, named host/stack/full settings, witness/tolerances and background-load method. No freeze, compute or acceptance authorized.
 
 [Campaign evidence](https://github.com/PyAutoLabs/autolens_profiling/pull/416)
 
 ### Active tasks
 
 - [Freeze the setup baseline specification, then collect and review evidence](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/setup_baseline_campaign.md) — needs-decision: HST/MGE raw fixed-parameter CPU runner merged in #416; Pulse handoff #55 merged and Mind implementation closed. Resolve exact production host/stack/data, full settings, independent correctness witness/tolerances and background-load method. No collection, campaign freeze or acceptance authorized; global compile dispatch remains unavailable.
-- [Prepare and review the HST/MGE fp64 CPU pilot](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/hst_mge_cpu_pilot.md) — needs-decision: HST/MGE raw fixed-parameter CPU runner merged in #416; Pulse handoff #55 merged and Mind implementation closed. Resolve exact production host/stack/data, full settings, independent correctness witness/tolerances and background-load method. No collection, campaign freeze or acceptance authorized; global compile dispatch remains unavailable.
+- [Prepare and review the HST/MGE fp64 CPU pilot](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/hst_mge_cpu_pilot.md) — needs-decision: Review the Pulse collection choices (candidate input hashes and 31-worker resource plan verified without execution) alongside project correctness audit #417. Resolve data origin/suitability, named host/stack/full settings, witness/tolerances and background-load method. No freeze, compute or acceptance authorized.
 
 </details>
 

tasks/hst_mge_cpu_pilot.md

@@ -199,3 +199,15 @@ Reviewed at 2026-10-13T17:51:40Z. HST/MGE raw fixed-parameter CPU runner merged
 GitHub sources and limits. Current RAL queue was empty; the transfer preview timed
 out, so historical jobs and unpulled results remain unverified. Review time does
 not refresh measurements or authorize compute. Historical task text is retained.
+
+
+## Collection choices prepared — 2026-10-13
+
+[Collection decision document](hst_mge_cpu_pilot_choices.md) records candidate
+FITS identities and the existing driver's plan-only resource bounds, distinguishing
+verified facts from unresolved production choices. The human requested this
+Pulse-side preparation under `--auto`, leaving the independent correctness audit
+[project #417](https://github.com/PyAutoLabs/autolens_profiling/issues/417) with its
+current session. No witness or profiling execution, campaign freeze or acceptance
+was authorized. This focused preparation does not advance the all-campaign
+`last_checkin` or measurement dates.

tasks/hst_mge_cpu_pilot_choices.md

@@ -0,0 +1,136 @@
+# HST/MGE CPU pilot — collection choices for review
+
+Campaign: setup-baseline
+Parent task: [hst_mge_cpu_pilot](hst_mge_cpu_pilot.md)
+Prepared: 2026-10-13
+Status: proposed choices; no production specification frozen
+Preparation issue: [Pulse #58](https://github.com/PyAutoLabs/PyAutoPulse/issues/58)
+
+The human requested continued preparation under `--auto` and explicitly kept
+[project correctness audit #417](https://github.com/PyAutoLabs/autolens_profiling/issues/417)
+with its current session. This document supplies the Pulse-side collection
+choices; it does not replace that audit, implement a witness, approve tolerances
+or authorize collection. Earlier approvals of tooling and preparation remain valid.
+The existing campaign and task remain the canonical queue; this supporting page
+is not a new experiment or an additional task.
+
+## What this pilot would establish
+
+A qualified pilot would measure cold compilation, demonstrated persistent-cache
+reuse and steady execution for one fixed-parameter HST/MGE CPU/fp64 objective on
+one frozen host/stack/input configuration. It would quantify repeatability within
+that configuration, with ten independent cold processes and ten seeded warm
+processes. The steady runtime's ten calls occur in one process: they do not
+establish between-process runtime variability or general modelling throughput.
+
+The pilot would not establish improvement over historical flat-vector posterior
+measurements, which use an incompatible objective. Compilation and execution are
+separate costs; the seed cost is retained. A finite output, successful cache hit
+or faster timing does not establish scientific correctness or baseline acceptance.
+
+## Verified candidate inputs — not yet selected for collection
+
+Reviewed project revision:
+[`e4266a1355a860d2274a82722858349cd3521e46`](https://github.com/PyAutoLabs/autolens_profiling/tree/e4266a1355a860d2274a82722858349cd3521e46).
+The local files below match the bytes at that Git revision. Dimensions and BITPIX
+were read from the primary FITS headers; values, units, masks, PSF normalization,
+noise validity and data origin were not scientifically validated in this step.
+
+| File under `dataset/imaging/hst/` | Header shape (NAXIS2 × NAXIS1) | BITPIX | File bytes | SHA256 |
+|---|---|---|---|---|
+| [data.fits](https://github.com/PyAutoLabs/autolens_profiling/blob/e4266a1355a860d2274a82722858349cd3521e46/dataset/imaging/hst/data.fits) | 141 × 141 | -64 | 164160 | `e4e833a28e9c6f6fb023cd85c0d3dde74068720256bc77a240633500e5b04fac` |
+| [noise_map.fits](https://github.com/PyAutoLabs/autolens_profiling/blob/e4266a1355a860d2274a82722858349cd3521e46/dataset/imaging/hst/noise_map.fits) | 141 × 141 | -64 | 164160 | `315c6266a0acd3ac9c8da82fcd2b45a08697362481fd7a35072bab1658530ba3` |
+| [psf.fits](https://github.com/PyAutoLabs/autolens_profiling/blob/e4266a1355a860d2274a82722858349cd3521e46/dataset/imaging/hst/psf.fits) | 21 × 21 | -64 | 8640 | `8389ba0e8bf4a09b4a561cfc246181a2fb944ed396754cd0ff76fdd22910d303` |
+
+These identify the existing candidate reproducibly; they do not prove that it is
+observational or simulated, suitable for the scientific question, or the same
+bytes as the eventual RAL copy. An instrument label and simulator code nearby
+are not origin evidence. Production selection must record origin and reverify
+these hashes on the execution host. If observational, inspect contaminants and
+agree the mask with the human before collection. Do not simulate replacement
+inputs silently.
+
+The [builder](https://github.com/PyAutoLabs/autolens_profiling/blob/e4266a1355a860d2274a82722858349cd3521e46/scripts/misc/cell_builders/imaging_mge.py)
+and [manual workload](https://github.com/PyAutoLabs/autolens_profiling/blob/e4266a1355a860d2274a82722858349cd3521e46/scripts/misc/jax_compile/current.py)
+declare 0.05 arcsec pixels, a 3.5 arcsec circular mask, radial LP oversampling
+[4,2,2] at [0.3,0.6], two 20-Gaussian light components, fixed mass/shear and
+prior-median free parameters. These source declarations are not an audit of the
+actual effective solver settings, input leaf types or masked-pixel count.
+Resolving the latter belongs with the project qualification work.
+
+## Verified plan-only capability
+
+The existing [project driver](https://github.com/PyAutoLabs/autolens_profiling/blob/e4266a1355a860d2274a82722858349cd3521e46/scripts/misc/jax_compile/pilot.py)
+was invoked without `--run`, using the candidate directory, a new unused output
+name inside `output/hst_mge_pilot_v1/`, and one CPU from the local allowed affinity.
+It returned a plan successfully, created no output directory and launched no
+workers. The local CPU selector only satisfied the planner; it is not a chosen
+RAL host or production affinity. No profiling or witness calculation was run.
+
+The plan reports `imaging-mge-runtime-v1`, `accepted: false`, matching input hashes,
+31 processes, 1800 seconds per process, 55800 maximum worker seconds, three
+runtime warmups and ten runtime samples. Its explicit qualification blockers are:
+
+- Independent scientific correctness reference and reviewed tolerances absent.
+- Background load excluding campaign work is not measured.
+- Frozen host/stack/dataset/configuration and collection approval not verified.
+
+## Choices to resolve before collection
+
+| Choice | Verified now | Proposal / next decision | Still required |
+|---|---|---|---|
+| Input dataset | Candidate identities above | Start by qualifying these existing bytes to preserve continuity with the shipped builder; use a different dataset only through an explicit scope decision | Establish origin, suitability, value/PSF/noise checks and execution-host hashes; observational inspection if applicable |
+| Hardware | CPU-only; production CPU policy requires `ral` | One named RAL CPU host and fixed core/affinity, one numerical thread, one worker at a time | Actual available node/class, CPU model/topology, affinity, RAM, storage quota and collection window; no allocation requested here |
+| Software | Producer implementation pinned above | Select a mutually compatible set of clean full Git SHAs plus a reproducible environment lock on the chosen host | Python/JAX/jaxlib/numerical packages, build flags, OS and all five PyAuto library revisions; a package inventory alone is not a lock |
+| Model and solver settings | Builder declarations above | Retain the narrow current-runtime CPU/fp64/jit objective; explicitly capture every effective setting | Project audit of objective normalization, parameter and coefficient ordering, mask/oversampling and actual solver configuration; no defaults changed |
+| Correctness | No independent witness or accepted tolerances | Consume the design from #417 before choosing implementation and acceptance rules | Human scientific review of independence, objective parity, absolute/relative bounds and near-zero handling; a separately scoped implementation/execution phase |
+| Background load | Runner reports load averages that include its own work | Develop an interval-based method separating campaign CPU activity from other work on the selected allocation, and retain raw counters/time intervals | Validate attribution on that host, including shared-core/SMT activity and scheduler context; the existing proposed ≤0.1 per allocated CPU bound has no accepted measurement method yet |
+| Cache evidence | Runner matches target module/key between seed miss and warm hit | Retain isolated cold caches and separate seed/warm pairs, with identical workload/stack/flags | Verify host storage and exact JAX diagnostic compatibility during authorized collection; faster compile alone is insufficient |
+| Resource ceiling | Driver caps workers/time; scheduler RAM is external | Keep the existing 32-process / 16 CPU-hour overall draft ceiling, with 16 GiB RAM per process | Human resource approval and scheduler enforcement; correctness process feasibility must fit its separate allowance or the proposal must be revised |
+| Collection and acceptance | Neither authorized | Review one complete project specification before requesting a collection window | Explicit compute authorization; independent later scientific acceptance and any baseline promotion |
+
+The background-load row is a method requirement, not a working estimator. Do not
+subtract a guessed one CPU from a load average or silently call it background
+load. The shipped `--xla_cpu_multi_thread_eigen=false` child setting is part of
+the method to freeze and record, not evidence that it is an optimal global flag.
+Any changed flag, host, stack, data, model, precision or estimator needs a distinct
+configuration identity; it is not a directly comparable before/after result.
+
+## Resource accounting and stop rules
+
+| Leg | Processes | Hard cap per process | Maximum CPU-hours at one allocated CPU |
+|---|---:|---:|---:|
+| Cold compile | 10 | 1800 s | 5 |
+| Cache seed | 10 | 1800 s | 5 |
+| Warm compile | 10 | 1800 s | 5 |
+| Steady runtime | 1 | 1800 s | 0.5 |
+| Independent correctness, reserved but not implemented by this runner | 1 | 1800 s | 0.5 |
+| Total proposal | 32 | | 16 |
+
+The shipped runner's portion is 31 processes / 15.5 worker-hours. The overall
+16-hour proposal is a worst-case CPU budget, not an expected duration or a
+scheduler wall-time estimate: queue time, driver overhead and allocation idle time
+can add elapsed time. At a 16 GiB per-process request, the 32-worker cap corresponds
+to 256 GiB-hours of worker resource envelope; scheduler billing/reservation can
+include extra time. No memory requirement was measured in this preparation.
+
+Stop on timeout, OOM, wrong backend/precision/context, missing input or provenance,
+invalid affinity/load evidence, failed correctness, or absent/ambiguous target
+cache evidence. Keep failures and missing legs visible. There are no silent retries,
+cap increases or GPU fallback. The independent correctness process may need more
+than the reserved allowance; #417 must establish feasibility before a final budget.
+
+## Handoff and next bounded action
+
+The correctness session continues #417. Pulse should review its resulting design
+alongside this document, resolve the dataset and host/stack choices, and identify
+any missing project implementation (witness, resolved-settings capture or load
+instrumentation) as bounded Mind tasks. No such implementation is started by this
+page. Project repositories own executable specifications, drivers and measurements.
+
+The eventual approval packet should link one immutable project specification,
+input/source/config hashes, environment lock, selected hardware/allocation, load
+method, correctness design and implementation, tolerance decision and total resource
+ceiling. Until these are resolved, collection remains unauthorized and the broader
+baseline manifest stays draft. Preparing this page changes neither `last_checkin`
+nor measurement freshness and does not create a decision-history entry.