Severity: high (completeness-critic finding; highest-leverage missing artifact)
bun test writes marker drawers into the production palace; eval baselines decay with the living corpus; the recall-roundtrip premise guard can't run offline. Root cause: there is no staging/fixture palace — a docker-compose postgres+pgvector+AGE seeded with a frozen 1–5K-drawer corpus and stable drawer IDs. One artifact simultaneously fixes prod pollution, makes recall-roundtrip runnable offline, stabilizes paraphrase-probe comparability, and gives the #77/#88 AGE A/B a reproducible substrate.
Fix: add ops/test-palace/docker-compose.yml + a seed script with a frozen fixture corpus; point tests/eval at it via env (PALACE_DAEMON_URL); assert in CI that tests never touch prod.
Filed from the 2026-06-09 dream-team premise audit (24 agents, adversarially verified). Report: docs/audits/2026-06-09-premise-audit.md.
Severity: high (completeness-critic finding; highest-leverage missing artifact)
bun testwrites marker drawers into the production palace; eval baselines decay with the living corpus; the recall-roundtrip premise guard can't run offline. Root cause: there is no staging/fixture palace — a docker-compose postgres+pgvector+AGE seeded with a frozen 1–5K-drawer corpus and stable drawer IDs. One artifact simultaneously fixes prod pollution, makes recall-roundtrip runnable offline, stabilizes paraphrase-probe comparability, and gives the #77/#88 AGE A/B a reproducible substrate.Fix: add
ops/test-palace/docker-compose.yml+ a seed script with a frozen fixture corpus; point tests/eval at it via env (PALACE_DAEMON_URL); assert in CI that tests never touch prod.Filed from the 2026-06-09 dream-team premise audit (24 agents, adversarially verified). Report:
docs/audits/2026-06-09-premise-audit.md.