[static] Previous-artifact cache silently expires after 1 day, forcing full 16-semester rebuilds

[tiankaima]

Problem

The curriculum/Young incremental cache relies on downloading the previous run's github-pages artifact (.github/workflows/build.yml:33-62). But actions/upload-pages-artifact@v4 defaults retention-days to 1 (verified in the action's action.yml), so each artifact expires exactly 24h after upload. The daily cron (.github/workflows/build.yml:5, 0 0 * * *) actually starts with ±40 min scheduler jitter — observed starts range 01:06–01:46 UTC. Whenever run N+1 starts later than run N's upload time + 24h, the restore step finds the artifact expired and silently rebuilds everything: all 16 selected semesters, ~280 JW schedule-table/datum chunk POSTs, all catalog lesson/exam fetches.

Evidence this already happened

• Run 29463669719 (2026-07-16 01:15:33) logged at 01:15:50: Previous build artifact 29380812432 is unavailable. The 2026-07-15 artifact was created 01:12:49 and expired 01:12:43 — 3 minutes before the restore step ran.
• The currently published snapshot proves the full rebuild: every row in upstream_fetches has fetched_at >= 2026-07-16T01:17:45 (checked https://static.life-ustc.tiankaima.dev/life-ustc-static.sqlite, SELECT MIN(fetched_at), MAX(fetched_at) FROM upstream_fetches → 2026-07-16T01:17:45 … 2026-07-18T01:17:11). With a working cache, ended semesters keep their original fetch rows.
• gh api repos/Life-USTC/static/actions/artifacts shows every github-pages artifact with expires_at = created_at + 1 day.

A full rebuild is ~10× the normal upstream load against jw.ustc.edu.cn / catalog.ustc.edu.cn and makes a mid-run failure (no publish at all) much more likely. It also recurs unpredictably since GitHub does not guarantee scheduled-run start times.

Suggested fix

One line in the upload step (.github/workflows/build.yml:80-84):

      - name: Upload static files as artifact
        id: deployment
        uses: actions/upload-pages-artifact@v4
        with:
          path: build/
          retention-days: 7   # survive a failed/delayed day

Optionally also make the restore step fall back to older successful runs when the newest artifact is expired (currently it only tries .[0].databaseId and gives up), and log a loud warning (::warning::) when the cache is lost so the full rebuild is visible instead of silent.