B3: register autofit_inference as instance fit; gaussian_x3 campaign and wave-1 task

[Jammy2211]

Refs PyAutoLabs/autofit_inference#4

Phase B3 of the search-extensibility epic, Insight side: register autofit_inference as instance fit and add the campaign and the wave-1 task.

Changes

• registry.yaml: new row fit. It points at autofit_inference dashboard/summary.json with schema inference-summary@1, library_refs [PyAutoFit] and cortex_project autofit_inference.
• campaigns.yaml: new campaign gaussian-x3-search-benchmark and task gaussian_x3_search_wave1. Both are blocked: the pilot stopped at 223 of 520 runs under the B3 wrap-up ruling, and the deferred cells wait on A2 (PyAutoFit#1679) plus a wave on RAL partition ral. NSS waits on A3b. job_status stays unknown, and nothing here authorizes a compute submission.
• tasks/gaussian_x3_search_wave1.md: the task intent, plus a status section for the wrap-up. It records that the pilot ranks nothing, that it ran on PyAutoFit 0dbf258c4f5e before A2, that the BFGS/LBFGS non-finite results are not a judgement on those searches, and that the @2 freeze is deferred.
• The board was re-rendered and the fit receipt and snapshot committed. tests/test_registry.py now expects both instances.

Note on the witness

The registry reads autofit_inference main. Today that holds the B2 summary with 0 records, which is valid, so check --offline passes. The 223 pilot records appear on the board at the next refresh after autofit_inference#5 merges. That is why autofit_inference merges first. I validated the PR's summary under this reader locally: classify returns ok with 223 records.

Decisions taken

1. Wrap-up ruling (human, 2026-10-08): no more pilot compute. The campaign and task are recorded as blocked, not complete.
2. The registry row lands now rather than after the merge. The B2 summary is already published on autofit_inference main, which meets B1's rule that a row needs a published summary, and check --offline passes.
3. The autofit_assistant campaign page is not made. autofit_assistant gitignores dated wiki/project/ entries, because they are per-clone memory and the template never ships them. The task file now points the verdict commentary at the autofit_inference ledger (wiki/project/state.md). Where the assistant page should live needs a human decision.

Checks (local)

ruff check and format pass. pytest tests passes (183 tests). bin/pyauto-insight check --offline is OK.

Merge order

1. autofit_inference#5
2. this PR
3. PyAutoCortex#61 (row autofit_inference set to active)

There is no autofit_assistant PR (see decision 3).

🤖 Generated with Claude Code