[Jammy2211]

Summary

This is the companion of PyAutoLabs/autofit_inference#7. Merge it after that PR. It is the Insight side of the B3c fixes from astra review 06 (autofit_inference#6). There are no scientific libraries, no compute and no contract version change.

Per item: finding, fix, test

1. Freeze rule (review finding 2). The next field in campaigns.yaml told the next session to run the 297 deferred cells and then freeze. That contradicted the no-restart ruling and the rule that the freeze comes before any scored run.

• Fix: the campaign gaussian-x3-search-benchmark and the task gaussian_x3_search_wave1 (in both campaigns.yaml and tasks/gaussian_x3_search_wave1.md) now state the same rule as autofit_inference protocol §7:
  • gaussian_x3@2 is frozen from the existing 17-row calibration before any scored run;
  • the pilot is not restarted, and its deferred runs move to scored wave 2;
  • scored wave 2 keeps the A2 + A4 gate.
• The blockers line now names A4 as well. The updated field is set to 2026-10-14.

4. Combined warm-start cost on the board (review finding 4). The producer now exports warm_start, provider_wall_s and consumer_total_s as additive fields. For warm-start rows, total_s is the combined attempt cost. The producer does the adding, so the reader still sums no clocks.

• Fix: board.total_value() shows that total and adds the provider's share, e.g. 1150.158048399997 (incl. nautilus warm-start provider 1076.4696579999982). This appears in both the markdown and HTML tables. Producers without these fields render exactly as before. REFERENCE.md documents the additive fields.
• Test: test_warm_start_total_shows_combined_cost_and_names_the_provider. It failed before the fix.
• Contract decision: stays inference-summary@1. The contract allows additive fields, and the meaning of total_s is set by each record's own definition text. @2 already names the setups and prepared-problems contract, which the fit producer does not emit. The reader validates the producer's regenerated branch summary (223 records) as ('ok', []).

6. Snapshot refresh and acceptance check (review finding 6).

• Finding: at the review's pinned merge, the fit capture was the zero-record B2 capture ec35be4. This does not reproduce on current main. The push-triggered refresh f2810fe (2026-10-08 19:54) had already captured autofit_inference 900648b5c665 with 223 records, and the board reads 35/223 accepted.
• Fix: new tests/test_fit_acceptance.py. It requires that:
  • the committed receipt and snapshot agree on the producer commit;
  • the capture is never the pre-pilot B2 one;
  • the capture holds exactly 223 records flagged pilot.
    For the known B3 producer revision (0c87bf1) and the B3c one (162644e), it also pins the 35 accepted / 110 rejected / 78 not assessed counts. A later producer revision keeps only the 223-pilot-records invariant, so routine refreshes do not turn it red.
• Verified that the new test fails on the B2 capture (8eaea76, 0 records) and passes on the current capture.
• Refresh steps:
  • pyauto-insight fetch --instance fit --mind ../PyAutoMind --out . resolved autofit_inference main to 900648b5c665, with 223 records and status ok. Only refreshed_at changed in the receipt.
  • pyauto-insight board --offline --mind ../PyAutoMind --out . re-rendered the dashboard, and the generated files are committed.
• The refresh must be repeated after autofit_inference#7 merges. Ingest always resolves the producer's main (registry.BRANCH), and the receipt-writing path cannot be pointed at a branch commit. Until that PR merges, the snapshot cannot carry the new warm-start fields or producer revision 162644e5c41a. The scheduled or dispatched dashboard refresh picks it up, and the acceptance check already covers that revision.

Findings that did not reproduce

• Finding 6 (board shows zero pilot records): fixed on main by the 2026-10-08 refresh, before this task. The acceptance check is added anyway, and it fails on the old capture.

Tests and gates

• python -m pytest tests -q: 186 passed (main had 183; 3 new tests).
• ruff check . and ruff format --check .: clean.
• bin/pyauto-insight check --offline --mind ../PyAutoMind: check: OK. Lens is 56 records at de37acfe; fit is 223 records at 900648b5c665; the dashboard is current and the state is valid.

Commits

• 6a8c689 campaigns: one gaussian_x3@2 freeze rule
• 418d6b5 board: combined warm-start total with the provider share named
• f5f69ab tests: B3 acceptance check on the fit capture
• 04c8026 insight: refresh fit receipt and re-render the dashboard

🤖 Generated with Claude Code