[Jammy2211]

Summary

Search-extensibility B4a, step 5. This registers autofit_profiling as Pulse instance fit (profiling-summary@2) and adds a hermetic fixture of its catalogue. It also updates the autofit campaign and the autofit_profiling_bootstrap task ("B4a at PR; B4b next"), and files epic 1's ranked bottleneck table as a dated campaign check-in.

Merge only after PyAutoLabs/autofit_profiling#3 is merged and its Pages deploy is live (https://pyautolabs.github.io/autofit_profiling/ serving catalogue.json). Pulse check fails any registered row without a published summary. Today fit has no receipt: autofit_profiling main has no dashboard/catalogue.json yet, and every ingest reads main (REFERENCE.md "Registry"). Writing a receipt from the PR branch would be a fabricated main receipt, so I did not write one.

Before merging, after #3 is live, on this branch:
bin/pyauto-pulse board (online), then commit receipts/fit.json, snapshots/fit.json, dashboard.md, dashboard.html, badge.json and state.json. check --offline then goes green.
If main has moved, merge it, re-render with the explicit repo flag (bin/pyauto-pulse board --mind <PyAutoMind>), and commit.

Per step

• registry.yaml: row fit, with repo: autofit_profiling, summary_path: dashboard/catalogue.json, supported_schema: profiling-summary@2, dashboard_url: https://pyautolabs.github.io/autofit_profiling/, library_refs: [PyAutoFit] and cortex_project: null.

• Fixture: tests/fixtures/autofit_profiling_v2.json is the producer catalogue copied verbatim from autofit_profiling#3 head 9074c7c.

  • New test test_the_autofit_profiling_producer_feed_validates_through_the_shared_reader reads it through summary.classify and expects ok, with runtime and breakdown axes and inapplicable slots carrying record_id: null.
  • The conftest body map gains autofit_profiling, and the committed-registry test now expects [lens, fit].

• campaigns.yaml: campaign autofit and task autofit_profiling_bootstrap are now active, updated: 2026-10-14, with next step "B4a at PR; B4b next…" and evidence autofit_profiling#3. last_checkin is untouched because this is not a full sweep.

• tasks/autofit_profiling_bootstrap.md: a "Check-in 2026-10-14" entry recording the B4a deliverables, the reproduction result, and epic 1's ranked table:

  Rank	Component	Share of wall
  1	NN training	43.4 %
  2	nautilus checkpoint I/O	24.5 %
  3	bound geometry	20.4 %
  4	visualisation	4.1 %
  5	bookkeeping	2.3 %
  6	samples output	1.7 %
  7	likelihood	1.1 %
  8	other	1.0 %
  9	proposal	0.8 %

  The entry also includes the B2/B3 likelihood_share from autofit_inference dashboard/summary.json at 637713b2, links pinned to 9074c7c, and records that epic 2 is not opened.

• Board re-rendered with board --offline. It currently shows fit as unavailable (red) until the receipt is captured as described above.

Pulse check output (bin/pyauto-pulse check --offline, this branch, before #3 merges)

ok   campaigns: 12 campaigns, 28 tasks
ok   body map: /home/jammy/Code/PyAutoLabs-wt/search-ext-b4a-breakdown-exporter/PyAutoMind/repos.yaml
ok   registry: 2 instance(s) (lens→PyAutoLabs/autolens_profiling, fit→PyAutoLabs/autofit_profiling)
ok   lens: profiling-summary@2 valid, 112 records, 0 comparisons at e4266a1355a8 (committed receipts + snapshots)
FAIL fit: unavailable: no receipt yet (never fetched)
ok   dashboard: current with every receipt
ok   state: valid against the Brain contract (red, 2 projects · 112 records · 0 cached · 1 failed)
check: FAIL

The only failure is the expected missing fit receipt. Everything else passes: campaigns, body map, registry with 2 instances, lens, dashboard currency and the state contract.

Decisions taken

• Epic-1 table filed as a task check-in, not a decision record. decisions/README.md reserves decision records for human decisions ("An agent's proposed interpretation is not a human decision"), and this table is a producer measurement plus a ranking. The full page lives in autofit_profiling wiki/campaigns/search_fit_bottlenecks.md.
• No receipt or snapshot for fit in this PR. The reason is the main-only ingest rule above. Capture happens after fix(lint): exclude github.com blob pages from lychee (503 to non-browser fetches) #3 merges.
• The campaign and task moved to active (from ready), because B4a is in flight.

Tests

• ruff check . and ruff format --check . are clean.
• pytest -q tests: 207 passed, including 1 new test.
• bin/pyauto-pulse check --offline: FAIL only on fit: no receipt yet, as explained above.

🤖 Generated with Claude Code