ci(fork-release-smoke): wait out boot1 teardown before seeding the scratch DB

[claudegoogl-sudo]

What Changed

• Commit 2 (seed the paperclip database, not postgres): the app creates and migrates the embedded database named paperclip (proven by boot1 logs); the seed connected to postgres where no app tables exist.
• fork-release-smoke.yml seed step: after boot1's teardown, poll pg_ctl -D <db> status until the old embedded postmaster reports not-running, clear the stale postmaster.pid/postmaster.opts, then start the seed cluster with pg_ctl -w -t 60 start (fail with the pg-seed.log tail on error).
• Upload sandbox/pg-seed.log alongside the other smoke artifacts.

Why Needed

• Run 34239890154 failed in the seed phase: preflight's teardown stops the scratch Postgres with SIGTERM and returns while the drain is still in flight (boot.log shows the SIGTERM at 14:48:02.2Z); the seed's pg_ctl start fired at 14:48:02.25Z against a live data dir → pg_ctl: another server might be running → could not start server.
• Install + boot1 phases passed on that run (checksums/closure/exports PASS, 317 packages from the release URL graph, GET / 200).

Verification

• Re-dispatch of the smoke workflow from this branch: GREEN — run 34241843769 (install from published URLs → boot1 migrations 0000..0249 → seeded stranded wake → boot2 → sweep resolved it, startup stranded deferred-wake sweep promoted or resolved wakes {scanned:1, resolved:1}).
• Artifacts from the failed run (install/boot1 evidence): https://github.com/claudegoogl-sudo/paperclip/actions/runs/34239890154

Risk Assessment

• Workflow-only change; release tree and tag untouched. Failure mode shifts from race to deterministic wait (60 s cap + explicit failure with logs).

Rollback Plan

• Revert this commit; the smoke workflow's previous behavior returns.

Migration/Compatibility Impact

• None (CI workflow only).

Model Used

• Paperclip prime_local adapter (provider-routed model; session does not expose exact backing model id)