Phase 1 — Single-node functional

Per-droplet smoke test. The release candidate's binary has to handle CLI roundtrips, MCP stdio handshake, curator dry-run, and backup on a fresh droplet before anything downstream is worth measuring.

Script: scripts/phase1_functional.sh.

Assertions (all must hold per node)

• ai-memory stats --json reports total >= 1 after the seed store.
• ai-memory recall <seed query> --limit 5 --json returns at least one hit.
• ai-memory backup --to <dir> --keep 1 leaves exactly one snapshot + one manifest file.
• ai-memory curator --once --dry-run --max-ops 10 --json reports at most one error (the "no LLM" case is accepted — Ollama isn't on the test droplets).
• ai-memory mcp responds to initialize + tools/list JSON-RPC frames with at least 30 tools in the response.

What failure looks like

Any assertion failing sets pass: false with a reasons: [...] list pointing at the specific check. Common causes and fixes:

Failure	Likely cause	Fix
stats.total<1	cloud-init finished but the write command errored (permissions, disk full)	df -h on the droplet; check cloud-init log
mcp.tools<30	Binary built without semantic feature OR integration test assertion needs update in ai-memory-mcp	Rebuild + rerun
snapshot!=1	Wrong --keep value or prior campaign debris left in /tmp/ship-gate-backup-<host>	rm -rf /tmp/ship-gate-backup-* and rerun
curator.errors>1	Non-LLM error from the curator — likely a code regression	Check the curator report's errors field

Output artifact

runs/<campaign-id>/phase1-node-<a|b|c>.json — one file per peer.