Step 5a sibling arm: synced sibling tests run against branch-era src/; #2208 collection probe cannot see the runtime skew

kind: infra. Workflow-surface gap in .claude/skills/issue/SKILL.md § Step 5a, the #1972 sibling-issue file freshness arm. Reproduced live on #2303, 2026-08-14, where it cost a full ~1h26m Step 9c gate run plus a re-gate.

Distinct from #2308 (Step 9c 1b newline splice) and from #2302 (the sibling arm's #2024 ordering_suspect carve-out). Same arm as #2302, different failure mode.

The defect

The sibling arm syncs a sibling issue's scripts/issue<M>_* and tests/test_issue<M>_* from origin/main into the current worktree, as a PAIR, so a gated test does not read a fork-era sibling script. src/ is deliberately not in SPECS — but a sibling's tests routinely import from src/explore_persona_space/experiments/issue<M>/.

So the sync can pull a sibling's current-main test into a worktree whose src/ is still at the branch's fork point. The test then exercises an API the worktree's module does not have yet, and fails — through no fault of the branch under review.

Concrete instance on #2303 (worktree forked ~2 h before the sync; the sibling's context_end work landed on main in between):

tests/test_issue2225_steer_hook.py:268
    mask = masks_for_mode("context_end", attention_mask=am, labels=labels)
src/explore_persona_space/experiments/issue2225/steer_train.py:81
    raise ValueError(f"unknown mask mode {mode!r}; expected one of {MASK_MODES}")
E   ValueError: unknown mask mode 'context_end'; expected one of ('all', 'context', 'response', 'prefix')

Two nodes failed this way. The Step 9c compare correctly classified both as NEW, so the gate blocked — on breakage the branch neither introduced nor touched.

Why #2208 does not cover it

The #2208 import-satisfiability probe reverts a synced issue-<M> pair when the synced test fails pytest collection in this worktree (the #2206 branch-era-src-import-skew shape: ImportError at collection time). That is the right guard for a missing symbol at import.

This failure is one layer later: the module imports fine, the test collects fine, and the mismatch only surfaces when a collected test calls an API whose behavior changed. masks_for_mode exists in both revisions; only its accepted argument set differs. A collection-time probe is structurally incapable of catching an argument/behavior-level skew, so the pair is never reverted and the gate reds.

Options (implementer picks; 1 and 2 are the cheap ones)

1. Extend the probe from collect-only to a bounded run. After the existing collection probe passes, run the synced sibling test file (short per-file timeout, -q --tb=no); on failure apply the SAME existing revert arm (restore branch-era content, or git rm a main-NEW file). Reuses #2208's machinery — only the probe's verdict source changes. Cost: the sibling test files' own runtime, on sync rather than at the gate. Guard against a sibling suite that is slow or needs GPU/network — cap it and treat a timeout as "leave synced, flag", never as a silent revert.
2. Sync the sibling's src/ subtree with the pair. Add src/explore_persona_space/experiments/issue<M>/ to what the sibling arm moves, making it a triple. Closer to the arm's own pair-atomic doctrine, but widens the blast radius into src/ — which SPECS has deliberately excluded, so this needs an explicit argument before adoption.
3. Classify it, don't prevent it. Have the Step 9c compare recognize a failure whose test file arrived via a sibling-sync commit AND whose traceback terminates in src/ as a distinct sibling-sync-src-skew class — reported, not counted NEW. Weakest option: the gate wall is still spent and a real regression could hide behind the label.

Whatever is chosen, the fallback the #2303 session used by hand should be documented in the arm's prose: bring the branch current (merge origin/main into the issue branch — not a rebase, which would need a force-push) so src/ and the synced sibling tests agree. That is the root-cause fix; reverting the pair only hides it.

Acceptance criteria

1. A worktree forked before a sibling's src/ change, whose sibling tests are then synced from current main, does NOT produce NEW gate failures attributable solely to that skew.
2. Whatever guard is added fails loud or reverts explicitly — never silently drops a sibling pair, and never silently strips a genuine failure.
3. A pin test reproduces the shape: a synced sibling test that COLLECTS cleanly but fails at runtime against branch-era src/. tests/test_issue_skill_lint_family_sync.py already hosts the sibling-arm _run_git harness and the #2206 repro, and is the natural home.
4. No change to the family arm (#2303's surface), the on-main skip, the #1972 selection logic, or the explicit-pathspec commit discipline.

Provenance

Surfaced by the #2303 orchestrator, 2026-08-14. Gate round 1: 3 failed, 8112 passed, 12 skipped in 5130.41 s; compare 2 NEW, 1 stripped. Resolved for that branch by merging origin/main (commit 642a792d39), after which the file passed 14/14 — but nothing in the repo was changed, so the arm still carries this hole.