fix(pg,gel): forward the selection argument through withReplicas $with

[pri12ya871]

What this fixes

$with takes (alias, selection?), but the withReplicas wrapper forwarded only the first argument:

const $with: Q['$with'] = (arg: any) => getReplica(replicas).$with(arg) as any;

Every other method on that wrapper spreads its arguments — select, selectDistinct, $count, with — $with alone did not. So the documented "CTE from raw SQL with a declared selection" pattern silently lost its selection under withReplicas, and every field read off the CTE came back undefined.

This is the first of the three problems @shadoworion reported in #6187. I have deliberately left the other two alone — with recursive support and array parsing for unbound sql fields are separate concerns with their own design questions, and folding them in here would make a one-line fix hard to review. Happy to look at either separately.

Scope

pg-core is where it was reported. gel-core carries the identical line and is fixed alongside it. The mysql, sqlite and singlestore wrappers name a $with variable but forward with, so they are a different shape and untouched here.

Test

integration-tests/tests/replicas/postgres.test.ts gains a test in the existing drizzle.mock() style, so it needs no database. It asserts three things: that the replica's $with is called with both arguments, that a field on the resulting CTE is defined rather than undefined, and that the generated SQL is right.

I confirmed it actually catches the bug rather than merely passing — with the fix reverted it fails with:

expected "$with" to be called with arguments: [ 'sq', { userId: PgSerial{ … } } ]
Received: 'sq'

and the file is 36/36 green with the fix.

One note on how I ran it: drizzle-orm's build script does not complete on my machine (Windows — scripts/build.ts exits during the zx-driven steps), and pnpm install at the root fails on a drizzle-kit@0.25.0-b1faa33 devDependency that is no longer published. I installed with --filter drizzle-orm --filter integration-tests and ran the replica tests against drizzle-orm/src through a local-only vitest alias, which is not part of this PR. So the test is verified, but by that route rather than the standard one — worth a second pair of eyes on CI.