Research runs here start with research/_shared/pick_track.py. Its first block, LAND FIRST, lists open pull requests titled research/... that are more than three hours old, so a run finishes old work before starting new work.
This morning two runs, 10:51 and 10:57 UTC, wrote the same complaint. The list showed #16494 as open and unlanded. Its content was already on main as a squash commit whose subject ended in (#16494). The pull request had stayed open with a merge conflict, so each run tried to merge it, hit the conflict and then worked out by hand that there was nothing left to land.
GitHub's list of open pull requests can't tell these cases apart. The commit history on main can. A squash merge ends its subject with (#N), and a merge commit starts with Merge pull request #N.
The change. LAND FIRST now reads the subjects of main's recent commits and checks the stale pull requests against them. It reads the local origin/main log and up to 300 more subjects from the public API, because a cloud clone is often only one commit deep. A pull request found there goes into its own group, "already on main: close it, nothing to merge", and drops out of the land-first count. On a live run the lookup read 350 subjects and matched the two newest merges.
The lookup is best effort. If the API can't be reached, the list falls back to its old behaviour and nothing is hidden. Two tests cover it: one for the squash and merge-commit subject forms, and one for the grouping. The other 59 tests in the file still pass.