test: make upstream PostgreSQL regressions primary - #204

Summary

Make the checksum-pinned PostgreSQL 18.4 regression suite the primary SQL regression suite while retaining architecture-specific integration, lifecycle, SDK, and packaging checks. Run unchanged upstream SQL, expected output, and result mappings through PostgreSQL’s own pg_regress.

• Run the full 231-case schedule against disposable native Oliphaunt clusters on Linux/macOS.
• Run a shared 29-case schedule against the actual embedded WASIX backend and TCP proxy, and extend WASIX Postmaster qualification with that schedule plus its existing upstream COPY case. Cases requiring host fixtures or the native regression C module remain covered by the native schedule.
• Wire the suites into Moon and existing CI lanes, preserve failure diagnostics, and document the testing policy. Add focused runner and CI-selection checks.

Release Intent

• Docs/CI/repository-only change: no release intended. No runtime implementation, package API, source pin, or generated runtime asset changes.

Verification

• Native upstream suite: 231/231 passed on Linux using available prebuilt PostgreSQL 18.4 Oliphaunt runtime artifacts.
• Shared embedded schedule against stock pinned PostgreSQL: 29/29 passed.
• Embedded WASIX: 28/29 upstream cases passed; all 7 existing architecture-specific tests passed, using the current binding code with available prebuilt runtime assets.
• Passed the runner behavioral test, existing Postmaster suite-name checks, shell syntax/ShellCheck, Rust formatting, documentation/tooling policy checks, release graph tests, and the complete bash tools/policy/check-workflows.sh gate, including the new affected-lane selection test.
• Postmaster execution was blocked locally by the missing patched Wasmer binary. macOS, Windows, and mobile were not replayed locally. These local results do not establish exact-commit hosted qualification.

Known failing gate

This PR is a draft because upstream json exposes an existing embedded WASIX session-recovery defect. After deliberate stack-depth errors with max_stack_depth = 100kB, RESET max_stack_depth and subsequent queries continue reporting stack-depth errors in that session. The test remains required and fails qualification; no expected-output override or ignored failure masks it. The runtime fix remains outstanding.