Expand integration test fixtures (schema_a/schema_b) to close diff-coverage gaps

[nettrash]

The data/test/schema_a.sql → data/test/schema_b.sql fixtures already cover the heavy machinery very well (partitions, persistence chains, virtual generated columns, CASCADE dependents, and every regression case from #180, #181, #186, #188, #191, #200, #202). However, a coverage audit against app/src/dump/ and app/src/comparer/core.rs surfaced a long tail of attribute-level diffs on otherwise-covered object kinds that are never exercised by CI.

CI runs the fixtures end-to-end against a real Postgres (.github/workflows/rust.yml), so each addition exercises the full dump→compare→apply pipeline.

High-priority gaps (verified absent via grep)

• GRANT privilege bits DELETE, TRUNCATE, REFERENCES, TRIGGER are never tested. Only SELECT, INSERT, UPDATE, EXECUTE, USAGE, CREATE appear in either schema. Add one table where schema_a has GRANT DELETE, TRUNCATE, REFERENCES, TRIGGER and schema_b drops one + adds another, so the diff exercises all four privilege-bit paths.
• No BRIN or HASH indexes. Only BTREE, GIN, and GIST are present. Add one BRIN index on an integer column (unchanged between A and B, to test dump round-trip) and one HASH index added in B (to test the create path).
• No functional / expression indexes. All indexes are simple column indexes. Add a functional index like (lower(name)) in schema_a that either changes expression or is dropped in schema_b, plus a new functional index added in schema_b.
• FK DEFERRABLE flag flip on a non-cycle FK is not tested. DEFERRABLE INITIALLY DEFERRED appears only in the cycle_a ↔ cycle_b pair (where it stays unchanged). Add a single non-cycle FK in schema_a without DEFERRABLE that becomes DEFERRABLE in schema_b — narrow but regression-prone area (recent Persistence-ordering silently emits arbitrary order on FK cycles #191).
• DEFAULT drop direction is not tested. customers.name gains a DEFAULT 'unknown', but no column has a DEFAULT in A that is removed in B. Add a column that has DEFAULT in schema_a and NO DEFAULT in schema_b.

Lower-priority gaps

• Index type change (e.g. BTREE → GIN) on the same column to exercise the drop-and-recreate path with a kind transition.
• UNLOGGED sequence flip independent of table ownership — tables flip LOGGED↔UNLOGGED but sequences do not in isolation.
• Function SUPPORT and TRANSFORM FOR TYPE clauses — Routine carries support_function and transform_types fields (app/src/dump/routine.rs) but no fixture exercises them. Narrow; rare in real schemas.

Out of scope (rejected as too narrow)

• Owner changes on indexes (PostgreSQL doesn't really support distinct index ownership)
• Overloaded function COST/ROWS where only one overload changes
• FK MATCH FULL / MATCH SIMPLE modifier flips
• View WITH CHECK OPTION / WITH LOCAL CHECK OPTION
• Operator operand-order asymmetry

Acceptance

When this is closed, every checked item above should have a matching A→B diff in the fixtures, the data/test/README.md matrix should mention it, and cargo run -- --config ../data/test.conf against the loaded databases should produce a clean migration script that, when applied to the A database, makes it schema-equal to B.

Total additions estimated at ~30–80 lines across both schema files.