[Pg-kit] Ignore leaf partitions when introspecting Postgres

[EndiButler]

What this fixes

A leaf partition has relkind = 'r', so Postgres introspection returns it as an ordinary table. Nothing downstream knows it is a partition.

Reproduced on this branch's parent (beta, drizzle-kit@1.0.0-rc.4) with one partitioned parent and three monthly partitions, against a schema file that does not declare them:

tables seen by introspection: [ 'logs', 'logs_2026_01', 'logs_2026_02', 'logs_2026_03' ]

push would run:
  DROP TABLE "logs";
  DROP TABLE "logs_2026_01";
  DROP TABLE "logs_2026_02";
  DROP TABLE "logs_2026_03";

Two separate problems:

1. push proposes dropping partitions the user never declared — for the common time-range retention setup, those are created by pg_partman or a cron job and hold the data the partitioning exists to protect.
2. It does not even complete: dropping the parent already dropped its partitions, so the second statement aborts with table "logs_2026_01" does not exist, leaving the database partially migrated.

pull is affected the same way — it emits a CREATE TABLE per partition, duplicating the parent's columns and its automatically propagated indexes.

This is what was reported in #2854 ("drizzle-kit push seems to be blowing up").

The fix

Filter on relispartition in both Postgres introspection paths (introspect.ts and aws-introspect.ts). The partitioned parent stays; its partitions are hidden. Because the child OIDs then drop out of every downstream constraint and index query, their auto-created indexes disappear from the pulled schema too.

Scope

Deliberately narrow: this is only the bug. It does not add any way to declare partitioning — that is the feature request in #2854, which I have written up separately as a design proposal with a working prototype branch. This fix stands on its own and does not depend on it.

Tests

• tests/postgres/pull.test.ts — the existing introspect partitioned tables test now creates a real leaf partition and asserts only the parent is returned.
• tests/postgres/pg-tables.test.ts — new push does not drop leaf partitions it does not know about: creates partitions outside drizzle, then asserts push produces no statements and both partitions survive.

Regression check on this machine: tests/postgres reports the same failure set as the untouched beta branch (those failures need a live Postgres/PostGIS container and a built drizzle-orm/dist), with one additional passing test. tsc --noEmit clean for both changed files.