When gh pr view Lies — The Stale-Base Branch Trap and a 33-Zombie Cleanup

The narrow miss

Two draft PRs opened earlier in the day looked routine:

• PR #484 — chore(daily-digest): framework cycle outputs 2026-05-25
• PR #488 — chore(integrity): stale-state-sweep run 2026-05-25 — noop_no_candidates

Both showed all CI checks green. Both had clean per-commit content. The gh pr view --json files API reported 2 and 6 files changed respectively. The PR descriptions matched what the operator intended.

The mergeStateStatus on both was BEHIND. That was the only public warning.

Then a closer look at git diff origin/main..<branch> --shortstat:

Merging either would have reverted ~10 of the same day's earlier merged PRs — including the HADF Sub-exp 1A key acquisition runbook, the OldSSD ↔ DevSSD migration verification addendum, the .env.local template the cloud-AI campaign was reading from, and all of the R-track lint config files (SwiftLint + ruff + markdownlint).

Why this happens

GitHub computes a PR's --json files against the branch's original base SHA, not against current origin/main. When main advances via squash-merges of unrelated work, every file added to main after the branch forked appears as "DELETED" in the merge-time effective diff. The branch isn't actively deleting them — it just doesn't have them, so merging as-is reverts them all.

The high merge velocity of the prior week amplified this: 7 PRs merged into main between 06:00 and 16:00 UTC the same day. Any branch that hadn't been rebased in those 10 hours accumulated phantom reverts.

The diagnostic that doesn't lie

git cherry-pick onto a fresh worktree off current origin/main. Cherry-pick uses patch-id matching — it can tell when a commit's content is already on main even if the SHA is different (e.g., the same change shipped via a different branch's squash-merge).

git worktree add -b _audit /tmp/audit origin/main
cd /tmp/audit
for c in $(git log origin/main..<branch> --reverse --format=%H); do
  git cherry-pick --keep-redundant-commits $c
done
git rev-list --count origin/main..HEAD     # 0 → ZOMBIE; >0 → has real content

Three outcomes:

What the audit at scale revealed

35 ahead-of-main local branches enumerated at session close. After applying the audit:

4   confirmed clean zombies   (cherry-pick → all empty)
11  bulk-audit zombies         (commits matched main via git cherry)
10  content-zombies            (conflicts revealed branch is older subset of main)
7   pr-* tracking branches     (all referenced CLOSED PRs)
1   claude/* session branch    (week-old, content shipped via main)
---
33  total deleted (94%)

The 2 survivors:

• 5 worktree-active branches — actively being developed (HADF, R9 coverage, analytics D3, UCC hardening, in-flight chore-green-items)
• 1 explicitly deferred branch — chore/dev-guide-v7-9-readability-pass-ft2-2026-05-24, intentionally paused per Mode B / Phase E hygiene

Genuinely live unmerged work: zero.

The validating meta-pattern

The team's session-close discipline of the prior 2-3 days had been so effective that most ahead-of-main branches in a high-merge-velocity repo are residue, not unfinished work. Each day's work cascaded into main via that day's PRs, leaving the local branch list to accumulate zombies.

This inverts the intuition. You'd assume an unpushed branch represents pending work — but in a healthy fast-moving repo, the more accurate prior is the opposite: assume zombie until proven live.

The fix is the diagnostic

Three operator actions, no new code:

1. At session close, enumerate ahead-of-main branches: git for-each-ref --format='%(refname:short)' refs/heads/
2. For each, run the cherry-pick audit (above)
3. Zombies + content-zombies → delete locally. Live → rebase + open fresh PR.

A 35-branch audit takes about 20 minutes. Skipping it saves nothing — the zombies will keep piling up, and the next near-miss will be a real merge instead of a draft.

Path 2: close + regenerate

For the two trapped drafts, the resolution was Path 2 — close the stale-base draft with an explanatory comment, then regenerate the intended change from scratch off current origin/main. PR #491 was the regen of #484: 3 files / +149 / -58 with no phantom reverts. The intended data update shipped cleanly within 30 minutes of the original draft closure.

Path 2 is the natural fix because the intended content is usually regenerable — daily-digest data is computed from current state via make documentation-debt + make measurement-adoption; integrity sweeps are computed fresh from current state.json files. The branch you'd rebase IS the branch you'd regenerate — so just regenerate.

What ships and what's queued

Shipped same session:

• W17 catalog entry (.claude/integrity/observed-patterns.md, PR #492) — full trigger / diagnosis / three-class taxonomy / silence path / examples / operator workflow
• PR #491 — the Path 2 regen of the closed #484 daily-digest
• 33 branches — deleted (zombies + content-zombies + closed PR tracking branches + stale session branches)
• 2 worktrees — pruned (123 MB reclaimed)
• 3 stashes — dropped

Queued as v7.9.1 candidate F-STALE-BASE-DETECTION:

A pre-merge gate extending the per-PR review bot — already worktree-captures origin/main baseline, so extending it to surface a P0 advisory when gh pr view --json files count vs git diff origin/main..<head> --stat count diverges by >2× is incremental work. Stacks with F-CONTRACT-FIXTURE-SAMPLING (W16) as the "PR-time sanity check" family — both close the silent-pass class of "the artifact you reviewed ≠ the artifact that ships."

Where this fits in the framework arc

W17 sits in a recurring family: the gap between what the agent sees and what the system ships.

• W9 (branch drift from concurrent-session collision) — agent thinks it's on branch A, another session checked out branch B in the same worktree
• W11 (incomplete PR cache) — agent reasons over a partial dataset, treats missing rows as absent rather than uncached
• W12 (vercel env pull returns empty for Sensitive vars) — agent reads env vars that look set but are silently undefined at runtime
• W16 (contract-boundary fixture sampling) — consumer tests pass against fixtures that encode the consumer's invented schema, not the producer's real one
• W17 (stale-base branches) — agent reviews a PR's file list that's correct against the branch's original base but misleading against current main

Each pattern looks unique on the surface but shares the same shape: a tool's output reflects a snapshot that's no longer the live state. The catalog discipline (CLAUDE.md v7.8.5 mandatory append-on-discovery rule) is the framework's response to this class. The next pattern in this family won't be the last.