[dan-petty]

Refine and the alternatives check wrote one result with two requests: the body, then the marks. A stop between them left the new body with the old marks. The next round read the new body as a person's edit and refined the item again, with model spend. The alternatives check looked unchecked and ran again. set_field and set_marks now take an optional body, and the GitHub store sends the body, the labels and the job record in one PATCH. Refine and the alternatives check no longer call write_issue_body for a result. Repair keeps it. The dry-run plan lists one request less. No setting changes.

Closes #4049

Change: openspec/changes/archive/2026-11-08-4049-one-write-body-and-marks/
Elapsed: 3.97s

[dan-petty]

Test record for #4049.

Fail-to-pass: uv run devops ci fail-to-pass --base origin/main ran in the tests stage. 13 tests failed on base and 86 passed. The 13 include all five criteria of the issue (refine write of a Ready item, a New item and a marks-only item; the alternatives write; the store body with marks). The tests are in tests/roadmap/refine/test_roadmap_refine_one_write.py, tests/roadmap/test_roadmap_alternatives_one_write.py, tests/roadmap/stores/test_roadmap_store_body_with_marks.py and tests/roadmap/test_roadmap_dry_runs.py.

Full uv run devops ci ran once, under the shared lock. One failure was mine: the alternatives rate-limit test injected its fault on write_issue_body. Fixed in f7db2b3 and the file passes. The coverage row failed only because the test row failed. The slow test row (13 minutes) is the known #3593. All other rows passed. After that run the branch was rebased onto origin/main 95f37d4 and the change was archived. The rebase touched only SearXNG tests and specs. The 66 tests of the five change test files pass again, and docs generate --check and openspec check pass. The full CI run was not repeated on the rebased tree; the CI checks of this pull request cover it.

Rollout rule: the change adds, removes, renames or tightens no setting. k8s/devops/configmap.example.yaml and the live ConfigMap are not touched, so no ConfigMap-load test is needed. Nothing was applied to the cluster.