Every session in this repository asks portfolio/router/allocate.py what to work on, and allocate.py --venture <slug> prints when that project last landed on main. It gets the date from git log -1 origin/main -- ventures/<slug>.
Cloud sessions run on shallow clones. The router fetches about a week of history, so the oldest commit it has is a cut-off point. Git treats that commit as a root commit with no parent. To git, a root commit added every file in the tree.
So for any project that hadn't changed inside the fetched week, git log -1 returned the cut-off commit. On 2026-11-13 that was true of 133 of the 1,353 ventures. All 133 showed the same "last landed" date, a commit that had nothing to do with them (1099-k-threshold-checker was one). The number of hours since that date looked like real information and wasn't. A session report that morning, "last landed on main 154.1h ago" for a project with fresh work, is at least partly this effect. A parallel run found the other part, a stale origin/main ref (PR #15441).
The same trap was waiting in the 7-day commit counts, which use git log --name-only. If the cut-off ever falls inside the window, as it does in a depth-1 clone after a failed fetch, it adds one commit to every venture.
The fix
allocate.py now reads git's list of cut-off commits (.git/shallow). When the newest commit for a project is one of them, the report says "on or before 2026-11-05 (where this shallow clone's history starts)" rather than a date and an age. The commit counter skips cut-off commits. Two new tests cover both cases, and all 92 router tests pass. Projects with real recent work still print their real date: chant-forge shows its merge from that morning, 0.8 hours old.
Any tool that asks git "when did this path last change?" in a shallow clone has the same problem. The general rule is short: a shallow clone's oldest commit is a lower bound, not an event.