Step 5a sibling-issue test sync can import a main-NEW test with branch-era-unsatisfiable src imports — deterministically reds the 9c gate as NEW (#2206 incident)

Bug

The /issue Step 5a sibling-issue file sync (.claude/skills/issue/SKILL.md, the #1972 per-FILE arm) can import a main-NEW tests/test_issue<M>_*.py file whose src/ import set is unsatisfiable in the branch-era worktree. The sync deliberately never touches src/ (correct), but it has no import-resolution check on the test file it imports — so the freshly-synced test deterministically fails COLLECTION (ImportError) in the Step 9c gate, and step9c_baseline.py compare classifies it NEW (fail-closed: the file IS touched by the branch diff via the sync commit, and the pristine oracle passes on main), blocking the gate on a failure that is not attributable to the round.

Incident (issue #2206, 2026-08-09)

• Step 5a sibling arm synced tests/test_issue2038_fallback_teardown.py (byte-identical to origin/main) into the issue-2206 worktree at commit 93d5161d2d.
• The test imports SupersededReapDecision from src/explore_persona_space/backends/issue_dispatch.py — a symbol added on main by the issue-2038 commit ed288f73f2, AFTER the worktree branch point; the worktree src copy is branch-era.
• 57m24s gate run: 8079 passed, 12 skipped, 1 collection ERROR (this file). Compare rc=1, new=[tests.test_issue2038_fallback_teardown], stripped/ordering_suspect empty.
• Resolved via the manual #542 provenance override (evidence in #2206's epm:test-verdict v1 marker). Cost: ~1h of gate wall + a compare cycle + manual evidence gathering.

Proposed fix (pick at implementation; (a) preferred)

(a) Import-check before sibling-sync of a test file. In the Step 5a sibling arm, after git checkout origin/main -- "$f" of a tests/test_issue*_*.py file, run a cheap collection probe in the worktree (uv run python -c "import ast; ..." static import scan against the worktree tree, or uv run pytest --collect-only -q "$f" with a short timeout); on failure, REVERT that one file (git checkout HEAD -- "$f" pre-commit / drop from SIBLING_SYNCED) and print the documented skip line (same shape as the existing "absent on origin/main" skip). Fail-safe direction: status-quo staleness, never an unreadable gate red.

(b) Alternatively/additionally: teach step9c_baseline.py compare to strip a NEW node whose file (i) was introduced by a commit whose subject carries the canonical sync workflow-surface specs from anchor phrase, (ii) is byte-identical to fetched origin/main, and (iii) passes the single-file pristine oracle — the same three probes the manual override used, mechanized.

Acceptance criteria

1. A worktree whose branch-era src lacks a symbol imported by a main-NEW sibling test no longer produces a blocking Step 9c NEW classification (either the file is not synced, or compare strips it) — pinned by a test reproducing the #2206 shape.
2. Legitimate sibling syncs (import-satisfiable test files) are unaffected.
3. The Step 5a SKILL.md prose + test_issue_skill_lint_family_sync.py-class pins updated together (family-atomic).

Files of record

Issue #2206 events.jsonl (epm:test-verdict v1, the two epm:progress diagnosis notes); .claude/skills/issue/SKILL.md § Step 5a sibling-issue file freshness (#1972); scripts/step9c_baseline.py compare classification; incident commits 93d5161d2d (sync), ed288f73f2 (main-side #2038 symbol).