Resolved: IS NULL/IS NOT NULL (#1), pg executor (#2), alias collision (#3).
New: Record DTS mismatch (#6), stock TS stack overflow (#7).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

ColumnInfo is the contract between introspectors and generateTable,
so it now speaks codegen's vocabulary — name / className / nullable /
default / generated — instead of information_schema-ese
(is_nullable: "YES", identity_generation, and the never-consulted
is_generated). Each introspector translates its dialect-native facts
at the boundary: PG maps identity + stored-generated columns; SQLite
maps notnull + rowid-alias detection, no longer pretending rowid is
IDENTITY "ALWAYS".

Behavior change hiding in the cleanup: PG `GENERATED ALWAYS AS (...)
STORED` columns previously emitted without `generated: true` because
only identity_generation was consulted and is_generated was dead
weight. The consolidated `generated` boolean covers both; locked by
a new name_upper column in the postgres e2e snapshot.

The sqlite/postgres e2e suites needed zero snapshot edits — the
payoff of pinning tests to DDL-in/code-out rather than the
intermediate shape this commit rewrites. Only generate.test.ts's
fixture helper changed. FkInfo left as-is (already dialect-neutral).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>