Symptom

A git audit of the kraken runner’s krill working copy found three ux-audit/* branches carrying real, never-reviewed UI token fixes — dated 2026-06-27, 2026-07-05 and 2026-07-06 — with no corresponding pull request on GitHub. The oldest had been sitting on the runner’s disk for over a month. Nothing surfaced this: the nightly audit reported success, main moved on ~250 commits, and the work was invisible to everyone.

Root cause

scripts/ux-audit.sh (kraken) creates its ux-audit/<date>-<theme> branch locally and only opens the PR at the end of the run. Any failure, abort, or non-zero exit between “branch created + committed” and “PR opened” leaves the commit stranded on the runner with no remote trace. Because the script also restores the working copy to its original branch on exit, the stranded branch is invisible in normal use — you only find it by enumerating local refs.

The same shape produced two duplicate copies of an old release-notes commit (kraken and release-notes-2026-06-27, both v1.0.1279) that likewise never reached a PR.

Fix

Recovered the two branches that still cherry-pick cleanly onto agents (2026-07-05 spacing, 2026-07-06 typography) into this PR; verified with :composeApp:compileKotlinDesktop. The 2026-06-27 branch and the stale release-notes commit conflict with current main and are reported separately rather than force-resolved a month after the fact.

Prevention

• A branch that exists only on the runner is not work — it is a leak. Any automation that commits to a branch must either open its PR or push the branch in the same run, so an interrupted run leaves a remote artifact someone can find.
• Prefer “push the branch first, open the PR second”: a pushed branch with no PR is recoverable and visible; a local-only commit is neither.
• Periodically audit long-lived runner working copies for local refs that are ahead of their remote — git for-each-ref --format='%(refname:short) %(upstream:track)' refs/heads makes the whole class visible in one command.