test(ai-persistence): static tests for the persistence adapter skills #1919

Problem

The persistence adapter skills in packages/ai-persistence/skills/ai-persistence/ have no tests. An agent following build-drizzle-adapter on Postgres wrote only chat_threads and chat_runs and dropped chat_interrupts and chat_metadata. PR #1918 fixed the skill text. Nothing stops the next regression.

Goal

Deterministic checks that run in pnpm test:pr and fail when a recipe is incomplete or does not compile. No LLM in the loop. Agent evals are a separate follow-up.

Tasks

1. Type-check the skill snippets with kiira. kiira.config already includes packages/*/skills/**/*.md, but the adapter blocks are marked ts ignore because they import @/db, @/db/schema, drizzle-orm, pg, and @prisma/client. Add those as docs-only dev deps and tsconfig.docs.json path stubs for the app-local imports, then drop ignore on the schema and store blocks in build-drizzle-adapter (SQLite and Postgres), build-prisma-adapter, and build-custom-adapter. Follow .agent/self-learning/lessons/2026-06-23-authoring-kiira-doc-snippets.md and kiira-ci-setup.md.
2. Run the conformance suite on the recipe code. Extract each recipe's chat-persistence.ts block into a vitest file under packages/ai-persistence/skills-test/ (or similar) and call runPersistenceConformance from @tanstack/ai-persistence/testkit against it. Drizzle: node:sqlite and PGlite (@electric-sql/pglite, new dev dep). Custom: PGlite via the pg-compatible driver. Prisma: skip if it needs a generated client; say so in the PR. Extraction can be a small script that pulls the fenced block by a marker comment so the test and the skill cannot drift.
3. Lint for all four tables. A small test that asserts each of build-drizzle-adapter, build-prisma-adapter, build-custom-adapter, and build-cloudflare-adapter names all four tables (threads, runs, interrupts, metadata) and all four store factories in defineAIPersistence.
4. Wire into an Nx target that test:pr already runs (test:lib in packages/ai-persistence is fine).

Out of scope

Agent evals (claude plugin eval, fixture apps, pass-rate thresholds). Track separately.

Related

• docs(ai-persistence): adapter skills always write all four chat tables #1918 (skill text fix)