Ban repo_root()-derived sys.path inserts in tests (worktree runs silently exercise main's code)

Goal

Add a workflow_lint.py check (plus its invariant test) that FAILS on any sys.path.insert(...) / sys.path.append(...) under tests/ whose argument is derived from task_workflow.repo_root(), and require the tree-local Path(__file__).resolve().parents[1] / ... form instead.

Why (concrete incident, task #2164)

tests/test_issue1482_densesae_fullwidth.py carried, at module level:

from explore_persona_space.task_workflow import repo_root
sys.path.insert(0, str(repo_root() / "scripts"))

repo_root() branch-guards to the MAIN checkout. Under any worktree-run pytest session that had two consequences, both silent:

1. The test exercised main's copy of the module under test, not the tree's. It imports issue1482_densesae_fullwidth from scripts/. From a worktree, repo_root()/scripts is the MAIN checkout — so a branch modifying that driver had its own test import a different file than the one it changed. A real regression in that driver could pass its own test on the branch and only surface after merge.
2. It leaked a foreign checkout's scripts/ onto sys.path for the whole pytest session. tests/test_backend_poll.py's #1296 negative control scrubs only entries resolving to the LOCAL tree's scripts/, so the leaked entry survived the scrub and a bare import runpod_api still resolved: Failed: DID NOT RAISE ModuleNotFoundError.

Effect (2) is how it was found: it turned #2164's Step 10d pre-push lint gate red as a payload-attributed NEW node, because that gate runs its mapped-test legs asymmetrically — gated from the worktree (where repo_root() does not match the cwd tree), baseline from the repo root (where it does, so the entry is scrubbed and the control passes). Diagnosis cost roughly ten probe commands plus a full ~20-minute gate re-run on a branch whose payload was innocent.

#2164 FIXED the single offending instance (landed on main via 5c91482fce; fix commit 69a58d6e5be439bc6465f767609267a75f784ad9) and swept tests/, scripts/, and src/ — it was the only occurrence at that time. This task adds the mechanical guard so it cannot come back, since nothing currently forbids it.

Scope

1. A workflow_lint.py check flagging sys.path.insert/append under tests/ whose argument expression mentions repo_root(. workflow_lint.py already carries an adjacent repo-root sys.path guard (see its docstring around the _ensure_repo_root_on_syspath() exemplar) — extend that family rather than adding a parallel mechanism, and bundle it into the no-flags default run so the Step 9c / Step 10d gates pick it up.
2. An invariant test for the new check, including a NEGATIVE control proving it fires on the pre-fix form (the deliberate-breakage non-vacuity pattern used by #2164's own anti-drift test).
3. Prefer monkeypatch.syspath_prepend over a bare module-level insert in the guidance text: tests/test_clean_experiment_downloads_symlinks.py:57 already documents that choice ("so the entry is restored"), so the repo has a sanctioned restoring form to point at.
4. Secondary, only if cheap: a one-line diagnostic note in the SKILL.md gate subsection saying that a NEW node appearing ONLY in the gated leg can be a worktree-vs-root environment artifact rather than a payload regression, with the two-file reproduction shape as the check. This is the diagnosis-cost half of the incident; the lint check above is the prevention half.

Out of scope

• Changing the gate's asymmetric baseline placement (gated $WT / baseline repo root). That asymmetry is deliberate per SKILL.md § Baseline semantics (the baseline must be a payload-free tree) and is not the defect here.
• The unrelated pre-existing test_setsid_child_survives_group_kill main-red, parked separately from #2164 in the #1681 urgent form.

Provenance

Surfaced by task #2164's Step 10d pre-push workflow-lint gate on 2026-08-07: verdict block on a NEW failing node that proved not to be payload-attributable (both files involved were byte-identical to origin/main on that branch).