[dan-petty]

Retitle took the same long-title items each run and made model calls for items it could not rename. It now records a skip mark (retitle.skipped, a hash of title, body and stage PR) for each item it leaves. It leaves the item out of the next selection until the title, body or stage PR change. This stops repeated model calls and lets later items reach the limit.

Three alternatives tests now read only the log records of their own logger. A span processor that an earlier test shut down added records of another logger, and the tests failed in one CI shard. This is part of the cause in #4062.

Closes #4038

Change: openspec/changes/archive/2026-11-07-4038-retitle-records-a-skip-mark/

[dan-petty]

Test record for #4038.

Fail-to-pass: uv run devops ci fail-to-pass --base origin/main (stage 2, at the test commit): 15 tests failed on the base, 2 passed (controls), 0 errored. File: tests/roadmap/test_roadmap_retitle_skip.py (17 tests).

Full uv run devops ci (before archive and rebase): exit 0, all rows pass. The Tests row took 21m59s, over the 5 minute budget. The slowest tests are tests/roadmap/stage/test_roadmap_stage_eval_*, the known class of #3593. This change did not cause it.

After the rebase onto origin/main 8476163: retitle_skip (17), retitle, dry_runs and stage_content tests (139 in total) pass. uv run devops docs generate --check passes.

Risk for review: the hash covers the whole item body, so a refine or alternatives block rewrite returns the item once to the selection. A model failure is recorded as a skip on purpose.