workflow-fix: single-flight repo-root sync helper (untracked-collision sweep)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from the 2026-07-02 shared-root divergence incident (PM-chat session, follow-on to the serial-compute audit batch #869-#876).

Goal

Add a single-flight root-sync helper so shared-repo-root reconciliation stops being an ad-hoc race between sessions.

Workflow gap

• Bug observed: on 2026-07-02 the repo root sat 582 local / 226 origin commits diverged for ~3h. Every pull-rebase recovery attempt (this session's and at least two other sessions' retry loops, observed live via pgrep) failed the same ways: (a) "untracked working tree files would be overwritten by checkout" — sessions continuously materialize eval_results/figures artifacts at the root uncommitted while worktree sessions commit the same files to origin (81 collisions in one attempt, 148 in the next, growing between attempts); (b) index.lock contention between concurrent recovery loops ("could not detach HEAD", "fatal: could not reset --hard"); (c) stranded/husk autostashes (three rescue stashes accumulated in one day). Origin/main had also LOST task #811's entire folder while its REGISTRY still pointed at tasks/interpreting/811 — recovered manually via a scratch-worktree merge (pushed as 6c1b3fadf7).
• Why it is a workflow gap: CLAUDE.md § Concurrent repo-root committers prescribes "git pull --rebase=merges --autostash once and re-push" but has no answer for untracked-collision failures, no single-flight coordination (N sessions all retry independently and collide), and no stranded-autostash recovery. The identical-vs-differing sweep that actually unblocks the pull was improvised inline twice today.
• Confidence (emitter): high (live incident; evidence in the 2026-07-02 PM-chat transcript).

Proposed change (candidate diff sketch — refine in planning)

New scripts/sync_repo_root.py (workflow-helper):

• Single-flight via flock on ~/.task-workflow/root-sync.lock (second caller exits 0 "another sync in flight").
• Pre-sweep: enumerate origin-only tree paths colliding with untracked root files; byte-identical → rm (content is in git); differing → move to a dated rescue dir (never delete non-identical data); report counts.
• git pull --rebase=merges --autostash with bounded timeout; on a mid-rebase conflict confined to tasks/<status>/<N> lifecycle files, auto-resolve keeping the LATER lifecycle state; anything else → abort cleanly + report (never leave a husk).
• Detect + restore stranded autostash entries; verify no MERGE_HEAD / rebase-merge husk remains.
• Push with one rebase-retry.
• CLAUDE.md § Concurrent repo-root committers: point the recovery sentence at the helper; session inline retry loops (the "for i in 1..5: pull" patterns observed today) call the helper instead.

Scope / surfaces

• Primary targets: new scripts/sync_repo_root.py, CLAUDE.md (§ Concurrent repo-root committers), tests
• Consider a watcher pass (autonomous_session_watch.py) invoking it when divergence exceeds a threshold — the plan decides whether that leg ships in v1.

Constraints / invariants

• NEVER a repo-root git reset --hard / git clean -f (standing ban); never delete non-identical untracked data (rescue dir only); fail loud, no silent husks.
• scripts/workflow_lint.py default run + ruff + tests pass.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 / a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: CLAUDE.md, scripts/sync_repo_root.py
• fingerprint: 0b509d58d95f

target_file: CLAUDE.md, scripts/sync_repo_root.py bug_observed: root pull-rebase recovery repeatedly fails on untracked eval_results/figures collisions materialized by concurrent sessions; multiple sessions race independent recovery loops (index.lock contention, could-not-detach-HEAD, stranded autostashes); local main sat 582/226 diverged for hours on 2026-07-02 why_workflow_gap: the documented recovery (pull --rebase=merges once) has no untracked-collision handling, no single-flight coordination, and no stranded-autostash recovery, so every session improvises and they collide proposed_change: add a single-flight scripts/sync_repo_root.py helper (flock; identical-rm / differing-rescue sweep of untracked collisions; pull --rebase=merges --autostash; push; stranded-autostash restore) and point the CLAUDE.md concurrent-committers recovery and session retry loops at it diff_sketch: |

• scripts/sync_repo_root.py: flock single-flight; pre-sweep untracked
• collisions (identical->rm, differing->rescue dir); bounded pull
• --rebase=merges --autostash; lifecycle-file auto-resolve; husk +
• stranded-autostash detection/repair; push with one retry
• CLAUDE.md: recovery sentence points at the helper confidence: high related_task: n/a (2026-07-02 divergence incident)