daily-fix: shared GCE staging HF-download retry

Overview / Motivation

Auto-filed by the /daily 2026-07-15 problem sweep (route 2 — behavior/logic change, independent review required).

Goal

Promote the hardened HF Hub download retry/backoff wrapper (built in-session on #1092) into the shared GCE staging path so per-issue scripts stop re-inventing it; consider defaulting the Xet/hf_transfer accelerators off for GCE staging phases with flaky egress.

Workflow gap

• Bug observed: #1092's rf01 GCE box crash-looped ≥3 same-day relaunches in the p6 static-staging phase on LocalEntryNotFoundError Hub connectivity; the session hand-built a hardened restore-and-retry wrapper + disabled xet/hf_transfer to get through (434a2458 19:05:37Z + att-190618/201123/202906).
• Why it is a workflow gap: the GCE lane's workload preamble/staging path has no shared hardened Hub-download retry, so every issue script that stages inputs on a GCE box either re-invents the wrapper after crashing or crash-loops through relaunches.
• Severity: medium
• verified-at-filing: grep -n 'LocalEntryNotFoundError\|retry.*download\|download.*retry\|hf_download' src/explore_persona_space/backends/gcp.py → 0 hits — no shared hardened download retry in the GCE lane (absence confirmed); the #1315-lineage transport-retry work (#1358/#1360 per the no-action inventory) targeted upload-policy/hub.py verify paths, not the GCE staging DOWNLOAD preamble (2026-07-16 UTC).

Proposed change (refine in planning)

Add a shared hardened Hub-download helper to the GCE staging path (workload preamble in src/explore_persona_space/backends/gcp.py, or an orchestrate/llm.hub staging helper the preamble sources): bounded retry with backoff around snapshot_download/hf_hub_download covering LocalEntryNotFoundError + connection-class failures, cache-restore between attempts, and an env knob to disable HF_XET_HIGH_PERFORMANCE/HF_HUB_ENABLE_HF_TRANSFER for staging phases on boxes with flaky egress (the #1092 in-session mitigation). Port the #1092 session's wrapper rather than re-deriving it; reconcile with the #1358/#1360 transport-retry helpers so there is one canonical retry surface.

Scope / surfaces

• Primary target: src/explore_persona_space/backends/gcp.py (workload preamble / staging)
• Secondary: the orchestrate staging helper if the wrapper lands as a shared module; .claude/rules/upload-policy.md cross-link (download-side sibling of the upload-wedge ladder)

Constraints / invariants

• Workflow-surface only (backends/*.py is in-scope workflow surface) — never experiment code, configs/, or tasks/.
• Fail-fast preserved: retries are bounded and the final failure surfaces loud (no silent fallback).
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

• fingerprint: f1e79e2f8f74

• workflow_fix_target: src/explore_persona_space/backends/gcp.py

Mined from 2026-07-15 session transcripts by the /daily problem sweep. Evidence: 434a2458 (#1092) 19:05:37Z + att-190618/201123/202906 (batch 07 P5).