RunPod SSH glob_sentinels — close #705's deferred stale-sentinel sibling probe over SSH

Overview / Motivation

Filed from #705's Step 5 reconciler verdict (round 1, PASS): the _default_glob_sentinels in src/explore_persona_space/backends/artifacts.py is local-FS only, so a RunPod live-pod multi-attempt stale-sentinel case (the #685 SECONDARY cause — declared sentinel missing, one live sibling exists in another attempt-dir under /workspace/eval_results/issue_<N>/) cannot be resolved by the new sibling probe over SSH. Codex's review correctly flagged this; the reconciler sustained the deferral on PASS because #685's primary cause ([git] not tracked by git) IS fixed and the secondary path FAILs honestly (no silent-pass regression), but the routine --skip-confirm-artifacts pressure for THAT specific RunPod stale-sentinel shape persists. This task closes the gap.

Goal

Wire an SSH-backed glob_sentinels for RunPod so the multi-attempt stale-sentinel sibling probe resolves over the live pod's filesystem too — matching the local-FS resolution #705 already shipped for GCP / SLURM.

Problem (from #705 review)

backends/runpod.py::RunPodBackend.confirm_artifacts injects read_sentinel=self._ssh_read_sentinel(handle) but NOT glob_sentinels. _resolve_live_sentinel therefore calls the default local-FS glob (Path(eval_results/issue_<N>).glob(...)) → zero local siblings → declared-missing FAILs honestly. A RunPod multi-attempt run with a stale-baked sidecar but one live sibling on the pod still forces --skip-confirm-artifacts.

Proposed change

• Add an SSH-backed glob_sentinels callable in backends/runpod.py that lists /workspace/eval_results/issue_<N>/*/.completion-sentinel.json over SSH via the existing SSH helper used by _ssh_read_sentinel.
• Inject it into the VerifierIO at RunPodBackend.confirm_artifacts next to read_sentinel.
• New regression test fixture mocking SSH ls + cat and asserting confirm_artifacts PASSes with exactly one live sibling, FAILs with zero, and FAILs with ≥2 (matching the local-FS contract in #705's tests).
• Same back-compat default: when glob_sentinels is None, behavior is unchanged.

Optional (named as standing concerns 2 + 3 from the same Codex review — defer to a separate task if they prove non-trivial):

• Scope-guard the glob to the canonical eval_results/issue_<N>/<attempt>/.completion-sentinel.json shape before resolving.
• Wrap the sibling liveness reads so an unreadable sibling returns a structured FAIL rather than escaping verify_artifacts.

Scope / target files

• src/explore_persona_space/backends/runpod.py
• src/explore_persona_space/backends/artifacts.py (optional — only if the scope-guard / read-error wrap is included)
• tests/

Constraints

• Workflow-surface only (in #705's in-scope list).
• Don't relax the gate: zero local siblings + zero remote siblings = declared FAIL, unchanged.
• Lint gate green; ruff clean on touched files.
• The local-FS path (#705 default) MUST stay unchanged — RunPod adds an SSH variant on top.

Provenance

• Parent: #705 (round-1 reconciler PASS verdict; sustained deferral framed by the plan v1 §12 row 6 Low-confidence assumption + implementer report's (b) Considered but not done).
• Linked open concerns:
  • runpod-ssh-stale-sentinel-deferred (#705, CONCERN, persisted by reconciler).
  • Standing recommendations from the same reconciler verdict: stale-sentinel-scope-guard-missing and stale-sentinel-sibling-read-raises (Codex CONCERNs).