fix(pgsql-test): fail the harness when seeding fails - #1774

Summary

getConnections caught every error from the seed phase, logged it to stderr, and handed the caller working clients on a half-built database:

} catch (error) {
  process.stderr.write(`[pgsql-test] Seed error (continuing):\n${formatted}\n`);
  // continue without teardown to allow caller-managed lifecycle
}

So a failed fixture deploy — the pgpm deploy that stands the schema up, or any seed.fn/seed.sqlfile — is not a setup failure but a passing setup, and the suite reports relation "…" does not exist in each test instead of the one error that caused it. Jest buries the stderr line above the first failure, which is how a single broken deploy reads as twenty unrelated mysteries.

Now the seed phase is fatal: stderr formatting is kept, the ephemeral database is torn down (its own failure deliberately swallowed with a comment, so it can't mask the seed error), and the original error is rethrown as the cause:

process.stderr.write(`[pgsql-test] Seed failed:\n${formatted}\n`);
try { await teardown(); } catch { /* abandoning this db anyway */ }
throw new Error(`[pgsql-test] Seed failed:\n${formatted}`, { cause: error });

Two tests cover it (a throwing adapter and invalid seed SQL). Measured before changing: no suite in this package, nor the db-heavy suites in constructive-db (metaschema, introspection, database-jobs, app provisioning), was riding on a swallowed seed failure — patched and unpatched runs are identical, and the rethrow was verified to actually fire with a deliberately-throwing adapter.

Link to Devin session: https://app.devin.ai/sessions/47477486a4fd45e684bd663b7007a629
Requested by: @pyramation