[Jammy2211]

Summary

B4b of the search-extensibility epic: ports the EP / graphical-model baselines from autofit_workspace_developer/{ep,graphical} and adds a profiling cell for the analytic Gaussian parity benchmark, reproduced on the B4a laptop host. The exporter reads two sibling result schemas and exports a runtime wall plus a memory host_peak_rss record per row on the existing profiling-summary@2 axes (no Pulse registry row, no contract bump). Findings: wiki/campaigns/ep_graphical_baselines.md.

Closes #4

Witness not met as written, for the human to judge: the witness asks for the reproduced N3 and N10 walls to be within runtime-drift-2x-1ms of the committed baseline.json. They are not. All six cells (EP and graphical, N3/N10/N30) are more than 2x faster (ratios 0.14–0.48, status improved). Correctness reproduces exactly: every truth logL sum matches bit-for-bit, and both N30 sanity FAILs reproduce. Nothing regressed. The committed baselines (2026-05-20) predate the changes behind the developer postfix re-runs, and the new rows are the baselines from here on.

Decision taken (flagged)

• Host pin: the epic's release-sweep pin (euclid-ral-gpu-2, A100) does not fit CPU Dynesty runs; B4b pins to the same laptop as B4a (DESKTOP-H143S82, WSL2, platform.node() in every row) so Pulse's same-host rule pairs these records with B4a's. Stated in the wiki page and the catalogue LIMITATIONS. Cross-machine reproduction on RAL is a later campaign.
• Drift notes are not Pulse comparisons: the committed developer baselines recorded no host, so no record pair exists; each row carries its runtime-drift-2x-1ms note in reference, and comparisons stays empty.
• Banked sigma_q95: the prompt says 12.18; the benchmark docstrings bank 12.15 and the closed form gives 12.1536. The cell checks against 12.15.

Reproduction (PyAutoFit df6d218f1, JAX CPU float64, one fresh-process fit per cell, serial)

method	N	wall s	committed s	ratio	drift	peak RSS MiB	truth logL sum exact	sanity (committed)
graphical	3	6.0	22.8	0.263	improved	420	yes	PASS (PASS)
graphical	10	7.1	50.7	0.141	improved	446	yes	PASS (PASS)
graphical	30	19.0	132.8	0.143	improved	515	yes	FAIL (FAIL)
EP	3	37.1	251.9	0.147	improved	353	yes	PASS (PASS)
EP	10	152.9	566.9	0.270	improved	403	yes	PASS (PASS)
EP	30	842.8	1758.7	0.479	improved	642	yes	FAIL (FAIL)

• Correctness reproduces: all truth logL sums equal the committed floats bit-for-bit; both N30 sanity FAILs reproduce (graphical: max-logL rule only, −1658.694 vs threshold −1650.561; EP: shared centre only, 49.26 ± 0.110, |z| 6.68).
• Witness not met as written: N3/N10 walls are not within 2x of baseline.json; all six are faster (improved). Against the developer postfix re-runs, graphical N3 is flat (1.06x), graphical N10 and EP N3 still faster (0.48x, 0.49x). No regression.
• Another session's CPU jobs shared the host (load 2.5 → 4.7), so the seconds are conservative; EP N30's N10→N30 scaling should wait for a quiet-host repeat.

Analytic parity (seed 0, N = 5): PARITY: PASS (41/41), both hard caps pass, wall 130.0 s, peak RSS 332 MiB, closed-form sigma_q95 12.1536. analytic_gaussian_priors.py is recorded as parked NEEDS_FIX, not run.

Scripts Changed

• scripts/gaussian_1d_ep/shared.py: shared simulator, sanity rules, harness and row writer, ported from the developer simulator/util/sanity.py.
• scripts/gaussian_1d_ep/{ep,graphical}/baseline.py: the EP and graphical N3/N10/N30 cells.
• scripts/gaussian_1d_ep/config/: verbatim developer config.
• scripts/analytic_gaussian/parity/benchmark.py: runs autofit_workspace_test's analytic_gaussian.py unchanged and records its verdicts.
• scripts/misc/tooling/build_catalogue.py: reads the autofit-profiling-baseline and -parity schemas; adds runtime + memory records, HTML sections and the host-pin and drift LIMITATIONS.
• scripts/misc/test/test_build_catalogue.py: 5 new tests (16 total).
• results/**, dashboard/**, wiki/campaigns/ep_graphical_baselines.md, README.md, AGENTS.md.

No PyAutoFit change.

Test Plan

• ruff check ., ruff format --check .
• pytest -q scripts/misc/test: 16 passed
• build_catalogue.py --check --validate-with ../PyAutoPulse: current, ok
• lychee: not installed locally (CI runs it)

Validation checklist (--auto run, plan was not pre-approved)

• Effective level: supervised (header: supervised, cap: feature/medium → safe)
• Plan: on the issue (feat: B4b EP / graphical baseline port + analytic benchmark cell (search-extensibility) #4)
• Gate: tests 16/16 · smoke n/a (profiling repo, no downstream script surface) · review: independent reviewer (fresh-context agent, not the author) returned FINDINGS ×2, both enacted in the second commit (wiki parity sha corrected to 26877bc69; rows carry provenance.host_contention and the catalogue names it in validation.reason; cells now record load averages). Its claim dispositions are all basis-cited: port faithful (side-by-side + diff -r config), truth logL bit-exact, N30 FAILs reproduced, ratios recomputed, deviations (a)–(d) disclosed honestly; Pulse re-ingest is pending sequencing · Heart STALE 85, acknowledged reason list: "release validation incomplete: no rehearsal for current source". STALE passes this leg. An earlier RED ("PyAutoFit: CI failure") was a timing flake in test_emcee_call_path_runs_at_jitted_speed on the 3.12 leg; the failed jobs were rerun green, then pyauto-heart tick
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject, then log the outcome

🤖 Generated with Claude Code