[finding] the new live-PostgreSQL cascade-delete cell is red-capable but CI never provisions it — the Temporal Conformance job runs driver-sql, not @objectstack/runtime #18734

Filed by the domain:engine execution seat (session_01CqmCgU5RGDoJYhHUMVp2af) out of the #18617 round (PR #18732), as the seat's ruling B on that round's open_questions[0]. ⛔ Filed bare: finding only; domain:* / type / priority are triage's.

Blocked-by: #18617

What exists after PR #18732, and what does not

PR #18732 gives packages/runtime/src/cascade-delete-multivalue-lookup-real-driver.integration.test.ts a driver axis: the embedded SQLite cell plus a live PostgreSQL cell, provisioned by OS_TEST_POSTGRES_URL. ⭐ The PostgreSQL cell is red-capable and was shown red — an ablation of SqlDriver.applyJsonMembership produced 6 failed | 9 passed, with all six failures in the live cell and zero in SQLite, reproducing operator does not exist: json ~~ text (42883) byte for byte.

⚠️ Nothing in CI hands that cell a URL. Measured by the filing seat on origin/main, ⛔ not carried over from the report:

⇒ That job provisions PostgreSQL and MySQL and runs driver-sql, the non-SQL temporal backends and metadata-protocol. It does ⛔ not run @objectstack/runtime, so the new cell reports itself un-run on every job.

What the cell does today without a URL — ⛔ it is not silent

The unprovisioned cell is a named skip that states which variable would run it, and under OS_EXPECT_LIVE_DIALECT_MATRIX=1 a missing URL is a failure rather than a skip, so a runner that claims to have provisioned a server cannot quietly degrade to SQLite-only. Measured: the full @objectstack/runtime suite without a URL reports 3661 passed | 1 skipped, and that one skip is this cell naming itself.

⇒ ⚠️ A named skip is a report, ⛔ not coverage. That is the gap this card carries.

Why this is its own card and ⛔ not part of PR #18732

The seat ruled B on the round's open question, and the reasoning belongs here too:

• The dispatch order told the dev to stop at .github/workflows/**, and it stopped. ⇒ Moving that fence because obeying it cost something would teach every later dev that the fence is negotiable.
• The leg is enforcement on a surface shared by every PR in the repo: it adds a step that can block anyone's merge. ⚠️ It is ⛔ not a new gate — it provisions an existing suite — but the blast radius is the whole merge path, and that deserves arguing in the open rather than riding in as the tail of a test PR.

⭐ And the distinction that makes landing #18732 safe rather than a punt: #18172 shipped because the PostgreSQL face could not go red at all — no cell existed. After #18732 a cell exists and has been shown red; what remains is provisioning. ⇒ A narrower defect with a different fix and a different owner — which is why it gets a carrier instead of a sentence in a PR body.

Shape (⛔ a proposal, not a prescription)

Two steps in the job that already provisions the postgres service, derived and written out by the #18732 dev:

      - name: Build runtime and its dependencies
        run: pnpm exec turbo run build --filter=@objectstack/runtime... --concurrency=4

      - name: Run the runtime cascade-delete matrix against live PostgreSQL
        env:
          OS_TEST_POSTGRES_URL: postgres://postgres:[email protected]:5432/postgres
          OS_EXPECT_LIVE_DIALECT_MATRIX: '1'
        run: |
          pnpm --filter @objectstack/runtime exec vitest run --project local \
            cascade-delete-multivalue-lookup-real-driver

⚠️ The positional is a SUBSTRING, ⛔ not a glob — the same reading the metadata-protocol step already records for itself, where the glob form matched zero files. ⇒ A taker that "improves" it into a glob will silently run nothing, which is the defect class this whole family is about.

⛔ Not measured

• Whether the taker should extend the existing job or add a new one, and what that does to job wall-clock on every PR. ⇒ A real cost this card ⛔ does not price.
• Whether OS_EXPECT_LIVE_DIALECT_MATRIX=1 belongs on the whole job or only this step.
• Whether any of the 29 other better-sqlite3 literals under packages/runtime/src/ should join the same leg. ⚠️ The test(runtime): pin the multi-value cascade-delete path on PostgreSQL, not only SQLite #18732 round deliberately did ⛔ not convert them (none carries a shipped dialect-specific defect), and this card ⛔ does not reopen that.

Dedupe words

OS_TEST_POSTGRES_URL runtime · cascade-delete matrix CI leg · live dialect matrix unprovisioned · Temporal Conformance runtime absent · named skip is not coverage

Refs: #18617 · PR #18732 (the pin and its ablation) · #18172 (the p1 that shipped in two releases) · #18200 (a different mechanism — which FILES a local green covers, ⛔ not which dialect)

Generated by Claude Code