workflow-fix: isolate root-guard self-test git state

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from the 2026-07-18 /daily transcript problem sweep (c3-P16). Route-2 filing.

Goal

Run the test_check_git_recipes_root_guard_* probe against an ISOLATED temp git repo (or serialize it on the shared-root lock) so a concurrent shared-root git op cannot make the live-hook probe read fail-open and produce spurious failures.

Workflow gap

• Bug observed: 13 spurious test_check_git_recipes_root_guard_* failures in #1506's gate — the guard probe read fail-open (blocked-probe rc=0) while a concurrent shared-root git op was in flight; cost a baseline compare + a 17-min full re-run.
• Why it is a workflow gap: check_git_recipes_root_guard invokes the live hook (guard_repo_root_branch.sh) as a subprocess probe; the hook's git-state read resolves against the AMBIENT shared repo rather than the tmp_path fixture, so a concurrent branch/reset op on the shared root transiently flips the probe's verdict — a fixture-isolated test that isn't actually git-isolated.
• Confidence (emitter): medium
• verified-at-filing: grep -n 'test_check_git_recipes_root_guard' tests/test_workflow_lint.py → the fixtures pass repo_root=tmp_path and write the recipe fence into tmp_path, but the invoked hook resolves git state in the ambient cwd (no per-probe GIT_DIR/cwd pin to the fixture); a concurrent shared-root git op therefore perturbs the probe (2026-07-19). NOTE the spawned session must first REPRODUCE the fail-open mechanism (run the probe under a simulated concurrent shared-root op) before choosing isolate-git-state vs lock-serialize.

Proposed change (candidate diff sketch — refine in planning)

# Pin the probe's git resolution to the fixture repo (isolate), e.g. run the
# hook with cwd=tmp_path + a tmp_path .git (git init), or GIT_DIR/GIT_WORK_TREE
# pointed at the fixture, so the ambient shared-root branch/HEAD is never read;
# OR serialize the probe on ~/.task-workflow/lock. Prefer isolation (removes
# the shared-state dependency entirely) over serialization (still contends).

Scope / surfaces

• Primary target: tests/test_workflow_lint.py, scripts/guard_repo_root_branch.sh
• Determine whether the fix belongs in the test harness (isolate the probe's git env) or in the hook (honor an explicit repo-root override); the test side is preferred if the hook already accepts a directory argument.

Constraints / invariants

• Workflow-surface only. Must not weaken the guard's real protection — the fix isolates the TEST probe's git state, not the live hook's shared-root guard.
• scripts/workflow_lint.py --check-asks passes; the full guard self-test matrix stays green AND becomes concurrency-robust.
• Recursion guard applies.

Provenance

• workflow_fix_target: tests/test_workflow_lint.py, scripts/guard_repo_root_branch.sh
• fingerprint: 722169522df4

Surfaced problem (c3-P16): 13 spurious root-guard test failures from a fail-open probe reading ambient shared-root git state during a concurrent git op; a baseline compare + 17-min re-run burned.