daily-fix: Step 10d merge contention in same-batch fleets

Overview / Motivation

Auto-filed by the /daily 2026-07-12 problem sweep (transcript-mined across 8 same-batch workflow-fix sessions: #1275/#1276/#1278/#1279/#1280/#1281/#1282/#1283).

Goal

Stop same-batch autonomous infra fleets from racing each other through Step 10d — cut the recurring ~30 min/night of merge-conflict recovery and redundant lint-gate re-runs.

Workflow gap

• Bug observed: on 2026-07-12, all 4 sessions of the 07:5x batch (#1275/#1276/#1278/#1279) failed their first gh pr merge --rebase ("Base branch was modified" / "can't be rebased" / real merge conflicts on tasks/REGISTRY.json + rename/rename on sibling task folders); 0/4 rebase-merges succeeded, all fell to --squash; #1278 needed a full scratch-worktree conflict recovery (~16 min). Same class the 07-11 daily counted in 10/24 sessions. Aggregate cost today ≈30 min wall + ~7 redundant lint-gate certification re-runs; a side effect was a transient duplicate tasks/planning/1280 folder (the #644/#1253 class, auto-remediated by the post-merge guard).
• Why it is a workflow gap: the watcher's proposed_infra_sweep dispatches up to the 5-session cap in one pass, so sibling infra sessions complete within minutes of each other; each one's status-move commits (git mv + REGISTRY rewrite) land on main between a sibling's lint certification and its merge attempt. Step 10d has no cross-session coordination, so the contention is guaranteed by construction, not incidental.
• Confidence (emitter): medium (the mechanism is confirmed; the best remedy is a design choice for the planner — merge-lock vs dispatch stagger vs squash-first default).
• verified-at-filing: transcript evidence 2026-07-12 sessions 004c5304/6fc58be1/aaf2482e/06281fbd (0/4 first-attempt merges; conflict quotes in #1278's transcript at 07:59–08:15Z); behavioral gap — not grep-verifiable as a single pattern.

Proposed change (candidate diff sketch — refine in planning)

Options for the planner to weigh (any one suffices):

1. A shared merge flock (e.g. ~/.task-workflow/step10d-merge.lock) held around the Step 10d certify→merge window, so sessions serialize the final stretch;
2. Stagger proposed_infra_sweep dispatches (e.g. one per watcher pass instead of cap-filling one pass);
3. Default Step 10d to --squash for autonomous infra sessions (rebase is 0/N under fleet churn anyway), keeping the certification SHA-bound retry ladder as-is.

Scope / surfaces

• Primary target: .claude/skills/issue/SKILL.md (Step 10d), scripts/autonomous_session_watch.py (proposed_infra_sweep dispatch pacing).
• Related shipped work: #1280 (Guard-1 three-dot scoping) removes the false-positive strip pressure but not the merge contention itself.

Constraints / invariants

• Workflow-surface only. scripts/workflow_lint.py --check-asks passes; ruff passes.
• Must not reintroduce repo-root branch work (the guard-blocked class) or force-push.
• This session runs under a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• fingerprint: d1856d16bb7a

• workflow_fix_target: .claude/skills/issue/SKILL.md

Origin: /daily 2026-07-12 transcript sweep (miner reports, sessions 004c5304/6fc58be1/aaf2482e/06281fbd + 3077c244/d78028e0/b49bb04a); 07-11 daily recorded the same class at 10/24 sessions.