feat: fit fixture + autofit campaign for B4b EP / graphical baselines

campaigns.yaml

@@ -80,11 +80,11 @@ campaigns:
   title: PyAutoFit profiling bootstrap
   status: active
   updated: '2026-10-14'
-  next: 'B4a at PR; B4b next. B4a = autofit_profiling#3 (search.fit breakdown cell, profiling-summary@2
-    exporter, Pages) then this Pulse fit row, then the autofit_workspace_developer retirement PR, merged in that
-    order. Epic-1 ranked table filed in tasks/autofit_profiling_bootstrap.md (2026-10-14); epic 2 not opened.
-    B4b = the EP / graphical baseline port.'
-  evidence: https://github.com/PyAutoLabs/autofit_profiling/pull/3
+  next: 'B4a merged (autofit_profiling#3, Pulse#60); B4b at PR (issue autofit_profiling#4) ports the EP /
+    graphical N3/N10/N30 baselines and the analytic parity cell (runtime + memory records, same laptop host as
+    B4a). Truth logL sums and both N30 sanity FAILs reproduce exactly; every wall is faster than the committed
+    developer baseline beyond 2x (improved). Receipt/snapshot refresh after its Pages deploy. Epic 2 not opened.'
+  evidence: https://github.com/PyAutoLabs/autofit_profiling/issues/4
 - id: critical-curves
   title: Critical curves and evaluation grids
   status: needs-decision
@@ -232,11 +232,11 @@ tasks:
   priority: normal
   path: tasks/autofit_profiling_bootstrap.md
   updated: '2026-10-14'
-  next: 'B4a at PR; B4b next. B4a = autofit_profiling#3 (search.fit breakdown cell, profiling-summary@2
-    exporter, Pages) then this Pulse fit row, then the autofit_workspace_developer retirement PR, merged in that
-    order. Epic-1 ranked table filed in tasks/autofit_profiling_bootstrap.md (2026-10-14); epic 2 not opened.
-    B4b = the EP / graphical baseline port.'
-  evidence: https://github.com/PyAutoLabs/autofit_profiling/pull/3
+  next: 'B4a merged (autofit_profiling#3, Pulse#60); B4b at PR (issue autofit_profiling#4) ports the EP /
+    graphical N3/N10/N30 baselines and the analytic parity cell (runtime + memory records, same laptop host as
+    B4a). Truth logL sums and both N30 sanity FAILs reproduce exactly; every wall is faster than the committed
+    developer baseline beyond 2x (improved). Receipt/snapshot refresh after its Pages deploy. Epic 2 not opened.'
+  evidence: https://github.com/PyAutoLabs/autofit_profiling/issues/4
 - id: cluster_pointsolver_speed
   campaign: cluster-pointsolver
   title: Cluster PointSolver — robustness, analysis settings and performance

dashboard.md

@@ -1,6 +1,6 @@
 # PyAutoPulse — profiling dashboard
 
-<!-- pulse-campaigns:40ebeabfcfa2bcc9dccff22b0816ae55956895694c0126147309ac3c6d3a7460 -->
+<!-- pulse-campaigns:b95c0bf384d49e4f8bd1f183ef71eae3e05fa2e4892d3ba474b05d4261563690 -->
 ## Profiling Check In
 
 <details><summary>Full check-in prompt</summary>
@@ -153,13 +153,13 @@ Reviewed 2026-10-13. Timing audit #410–#412 and raw HST/MGE pilot producer #41
 
 <details><summary>PyAutoFit profiling bootstrap — active</summary>
 
-Reviewed 2026-10-14. B4a at PR; B4b next. B4a = autofit_profiling#3 (search.fit breakdown cell, profiling-summary@2 exporter, Pages) then this Pulse fit row, then the autofit_workspace_developer retirement PR, merged in that order. Epic-1 ranked table filed in tasks/autofit_profiling_bootstrap.md (2026-10-14); epic 2 not opened. B4b = the EP / graphical baseline port.
+Reviewed 2026-10-14. B4a merged (autofit_profiling#3, Pulse#60); B4b at PR (issue autofit_profiling#4) ports the EP / graphical N3/N10/N30 baselines and the analytic parity cell (runtime + memory records, same laptop host as B4a). Truth logL sums and both N30 sanity FAILs reproduce exactly; every wall is faster than the committed developer baseline beyond 2x (improved). Receipt/snapshot refresh after its Pages deploy. Epic 2 not opened.
 
-[Campaign evidence](https://github.com/PyAutoLabs/autofit_profiling/pull/3)
+[Campaign evidence](https://github.com/PyAutoLabs/autofit_profiling/issues/4)
 
 ### Active tasks
 
-- [autofit_profiling: bootstrap the repo + general PyAutoFit profiling epic](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/autofit_profiling_bootstrap.md) — active: B4a at PR; B4b next. B4a = autofit_profiling#3 (search.fit breakdown cell, profiling-summary@2 exporter, Pages) then this Pulse fit row, then the autofit_workspace_developer retirement PR, merged in that order. Epic-1 ranked table filed in tasks/autofit_profiling_bootstrap.md (2026-10-14); epic 2 not opened. B4b = the EP / graphical baseline port.
+- [autofit_profiling: bootstrap the repo + general PyAutoFit profiling epic](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/autofit_profiling_bootstrap.md) — active: B4a merged (autofit_profiling#3, Pulse#60); B4b at PR (issue autofit_profiling#4) ports the EP / graphical N3/N10/N30 baselines and the analytic parity cell (runtime + memory records, same laptop host as B4a). Truth logL sums and both N30 sanity FAILs reproduce exactly; every wall is faster than the committed developer baseline beyond 2x (improved). Receipt/snapshot refresh after its Pages deploy. Epic 2 not opened.
 
 </details>
 

receipts/fit.json

@@ -13,5 +13,5 @@
   "record_count": 31,
   "errors": [],
   "cached_from": null,
-  "refreshed_at": "2026-10-15T12:22:48Z"
+  "refreshed_at": "2026-10-15T12:23:36Z"
 }

receipts/lens.json

@@ -13,5 +13,5 @@
   "record_count": 112,
   "errors": [],
   "cached_from": null,
-  "refreshed_at": "2026-10-15T12:22:47Z"
+  "refreshed_at": "2026-10-15T12:23:35Z"
 }

tasks/autofit_profiling_bootstrap.md

@@ -63,6 +63,15 @@ not refresh measurements or authorize compute. Historical task text is retained.
 
 - Next: B4b (EP / graphical baseline port). Epic 2 (EP loop) is **not** opened; it waits for epic 1's top findings to ship. Candidate one-PR prompts (checkpoint cadence, end-of-fit visualisation) are listed in the project page for the human / profiling conductor, not filed.
 
+
+## Check-in 2026-10-14 — B4b at PR; EP / graphical baselines reproduced
+
+- B4b (search-extensibility, issue autofit_profiling#4) ports `autofit_workspace_developer/{ep,graphical}` as `scripts/gaussian_1d_ep/{ep,graphical}/baseline.py` and adds `scripts/analytic_gaussian/parity/benchmark.py`, which runs autofit_workspace_test's analytic parity script unchanged. Each row exports a `runtime` wall and a `memory` `host_peak_rss` record on the existing axes (sibling result schemas `autofit-profiling-baseline` / `-parity`; no registry row, no contract bump). Findings: `wiki/campaigns/ep_graphical_baselines.md` in autofit_profiling.
+- Host pin (decided and flagged): the same laptop as B4a (`DESKTOP-H143S82`, WSL2), not the release-sweep A100 node, so Pulse's same-host rule pairs the new records with B4a's. PyAutoFit `df6d218f1`, JAX CPU float64.
+- Reproduction against the committed developer `baseline.json`: every `truth_log_likelihood_sum` matches exactly and both N30 sanity FAILs reproduce (graphical: max-logL rule; EP: shared centre |z| 6.7). Every wall is faster beyond the `runtime-drift-2x-1ms` policy (graphical 6.0 / 7.1 / 19.0 s vs 22.8 / 50.7 / 132.8 s; EP 37.1 / 152.9 / 842.8 s vs 251.9 / 566.9 / 1758.7 s; all `improved`), so the "N3/N10 within 2x" witness is not met as written; the committed baselines predate the changes behind the developer postfix re-runs. Another session shared the host during the runs (load 2.5–4.7), so the seconds are conservative.
+- Analytic parity (seed 0, N = 5): PARITY PASS 41/41 for closed form vs minimal EP vs graphical vs EP; closed-form `sigma_q95` 12.1536 matches the banked 12.15; `analytic_gaussian_priors.py` recorded as parked NEEDS_FIX, not run.
+- This Pulse branch refreshes the fixture; the `fit` receipt and snapshot are refreshed here after the autofit_profiling PR merges and its Pages deploy is live. Measurements are `unreviewed`; the Brain's profiling conductor judges.
+
 ---
 
 # autofit_profiling: bootstrap the repo + general PyAutoFit profiling epic