[Jammy2211]

Summary

Search-extensibility B4a: autofit_profiling gets its first measurement and becomes a PyAutoPulse producer. The Nautilus search.fit anatomy from autofit_workspace_developer/searches_minimal/ is re-homed as a breakdown cell, run on the laptop CPU, published as a thin profiling-summary@2 catalogue with a Pages board and a pulse-refresh dispatch, and epic 1's ranked bottleneck table is written up with likelihood_share from autofit_inference.

Closes #2. Merge order: this PR first, then PyAutoPulse (the fit registry row reads the published catalogue), then the autofit_workspace_developer deletion PR.

Per step

1. Contract read — Pulse REFERENCE.md v2 section, pulse/summary.py::classify, autolens_profiling build_catalogue.py (shape only) and pages_dashboard.yml, and the Mind record complete/2026/10/pulse-refresh-project-identity.md.

2. Breakdown cell — scripts/gaussian_x1/nautilus/search_fit_breakdown.py. Monkey-patches the nautilus internals (NeuralNetworkEmulator.train, NautilusBound.compute, Sampler.sample_shell/evaluate_likelihood/write/write_shell_update/run) and PyAutoFit's output path (SearchUpdater._save_samples/_compute_latent_samples/_profile_and_summarize/visualize, NonLinearSearch.pre_fit_output/post_fit_output). It charges exclusive wall time through a stack timer (scripts/misc/profiling/exclusive_timer.py), so nested calls are never double-counted and other = wall − instrumented. There are two variants:

   • search_fit: the full af.Nautilus(...).fit, written to results/gaussian_x1/nautilus/breakdown.json.
   • bare_sampler: the findings page's own bare-nautilus.Sampler measurement re-run as the reproduction gate, written to breakdown_bare_sampler.json. The cell exits 1 if NN training, bound geometry or likelihood fall outside ±10 pp of the findings page.

   Both use seed 1 and are deterministic in calls (7 501 and 16 400 likelihood calls in every repeat). Each is the median of 3 fresh fits. The findings prose is in wiki/campaigns/search_fit_breakdown.md, with a new §0 for the re-measurement.

3. Exporter — scripts/misc/tooling/build_catalogue.py (stdlib, about 300 lines). It turns results/**/*.json into dashboard/catalogue.json, index.html and state.json, with one setup per variant, one breakdown record per component plus a runtime wall record, and an explicit selection for each. A zero checkpoint_io (the bare sampler writes no checkpoint) becomes an inapplicable slot, not a zero record. Every record is unreviewed. Rendering is deterministic because the stamp is the newest measured_at. Use --check for staleness and --validate-with PULSE to run Pulse's own classify. Tests are in scripts/misc/test/. lint.yml now runs pytest plus build_catalogue.py --check --validate-with _pulse (PyAutoPulse checked out after ruff).

4. Pages + dispatch — .github/workflows/pages_dashboard.yml mirrors autolens_profiling's workflow. It validates the catalogue with Pulse and the badge feed with Brain board/_state.py, deploys dashboard/, then sends repository_dispatch: pulse-refresh (repo, sha, dashboard/catalogue.json) using PAT_PYAUTOLABS. Without the secret it warns and exits 0. state.json uses organ: autofit_profiling, following the project-feed identity in the Mind record.

5. (PyAutoPulse PR, separate.)

6. Epic-1 table — wiki/campaigns/search_fit_bottlenecks.md. Epic 2 is not opened.

7. (autofit_workspace_developer PR, separate.)

Measured breakdown (laptop CPU, DESKTOP-H143S82, WSL2 8 cores, numpy fp64, PyAutoFit 911fe05e1 = 2026.10.12.1+1, nautilus 1.0.5, n_live 200, n_networks 4, 1 core)

Reproduction (bare sampler): wall 22.5 s (repeats 23.1/22.5/21.8). log Z −57.455, the same as the findings reference.

component	findings §1	measured	Δ
NN training	51.7 %	53.8 %	+2.1 pp
bound geometry	25.2 %	25.3 %	+0.1 pp
likelihood	14.3 %	11.6 %	−2.7 pp
proposal sampling	5.2 %	5.7 %	+0.5 pp

All three gated shares are within ±10 pp, so the cell exits 0.

search.fit (epic 1): wall 28.4 s (25.8/28.4/28.9), 7 501 calls. Shares of wall:

component	share
NN training	43.4 %
checkpoint I/O (nautilus HDF5, 103 writes, about 69 ms each)	24.5 %
bound geometry	20.4 %
visualisation	4.1 %
sampler bookkeeping	2.3 %
samples output	1.7 %
likelihood	1.1 %
other	1.0 %
proposal sampling	0.8 %

likelihood_share cross-check: B3 Nautilus numpy has a median of 1.9 % (n=11), and Nautilus JAX CPU about 0.01 %.

Pages status

Pages is enabled (gh api repos/PyAutoLabs/autofit_profiling/pages -X POST -f build_type=workflow returned build_type: workflow, html_url: https://pyautolabs.github.io/autofit_profiling/). It is not live yet: the URL returns 404 until pages_dashboard.yml runs on main after this merges, because it triggers on dashboard/** pushes to main. I could not verify that the org secret PAT_PYAUTOLABS is visible to this new repo (reading org secrets needs admin:org). If it isn't, the deploy still succeeds and the dispatch step only warns. In that case, run Pulse's dashboard_refresh.yml by hand.

Validator output (tail)

pulse validator: ok (profiling-summary@2)
catalogue: current

Decisions taken

• Two variants, not one row. The findings page measured a bare nautilus.Sampler, while a PyAutoFit search.fit differs in three ways: it checkpoints (24.5 %), stops at n_eff=500 (7 501 calls against 16 400), and has a cheaper per-call likelihood (about 43 µs through Fitness against about 160 µs through the bare script's instance_from_vector). Its likelihood share (1.2 % sampler-internal) therefore cannot reproduce 14.3 %. The reproduction gate runs on the faithful bare-sampler variant, and the search_fit row records its comparison as information only (gates_reproduction: false).
• Tolerance ±10 pp absolute per share. The findings page states ±40 % session variance in wall time, and shares are the stable quantity.
• The fast 10-d scenario was not ported. It is not a PyAutoFit search.fit. Its numbers stay quoted on the findings page.
• likelihood_share source. catalogue/search_catalogue.json has no likelihood_share field on main 900648b5 or on feature/search-ext-b3-postmerge-fixes 637713b2 (it aggregates legs by task). The per-run values come from autofit_inference dashboard/summary.json (inference-summary@1) at 637713b253ab149f87a08ea0b02fcb009b4aa24e. The values are identical on main.
• Library output defaults. config/general.yaml only disables the workspace version check, so the cell measures what a user's search.fit pays. Visualisation is PyAutoFit's search-level default; there is no custom Analysis visualiser.
• Every record is validation: unreviewed, with qualified: true because provenance is complete. Pulse never judges, and the profiling conductor has not reviewed these numbers.
• producer_revision: null. The catalogue is committed alongside the rows it describes, so no self-referencing SHA is possible. generated_at is the newest measured_at, which keeps --check deterministic.
• The wall is a runtime record (metric: wall) and components are breakdown records. Shares are not catalogue records, because a breakdown unit must be s. They are shown on index.html and in the result rows.
• Project badge status: green with a neutral headline. It reads "baseline only, not a verdict", and the Brain cockpit reads the Pulse organ feed, not this one.

Tests

• pytest -q scripts/misc/test: 11 passed. The exclusive timer has 4 tests. The exporter has 7, including the committed catalogue being current, the committed rows validating with the Pulse reader, and the Pulse reader refusing a broken unit.
• ruff check . and ruff format --check . are clean.
• build_catalogue.py --check --validate-with ../PyAutoPulse reports ok.
• Brain board/_state.py dashboard/state.json reports state: ok.

🤖 Generated with Claude Code