The repository's local CI runner, platform/ci/run_workflow_locally.py, throws away its verdicts when the checkout changes under it. That guard exists for a real incident: on 2026-09-11 a sweep read another branch's tree halfway through and reported a stale mirror that was not stale.
A fix on 2026-11-13 taught the guard that a git commit of edited files changes no byte a step reads. It still had a gap. The fingerprint covered tracked files only, so a file that was untracked when the sweep started (a new blog post, the run record) and then got committed looked like a file appearing from nowhere. Two run records from the early hours of 2026-11-14 report five workflows voided this way, each a re-run of several minutes.
The fix takes one extra reading before the sweep: the blob ids of untracked, non-ignored files. A path that is tracked afterwards but was not before now counts as unmoved when its bytes match that untracked reading. Anything else still voids the sweep: a new file with different bytes, a file that was never in the worktree (a merge or a branch checkout), or an edit to a tracked file. Three new tests check those three cases against a real git repository, and the existing 171 tests still pass.
If the untracked set is over 2,000 files, or git cannot answer, the lookup is skipped and the guard behaves as before: stricter, never looser.