Use DBMS_METADATA for structure dump

[yahonda]

Summary

Adds a second structure_dump implementation built on Oracle's DBMS_METADATA.GET_DDL / GET_DEPENDENT_DDL alongside the existing data-dictionary (ALL_*) one. The new path lives in a separate file (lib/active_record/connection_adapters/oracle_enhanced/structure_dump/dbms_metadata.rb) — the existing structure_dump.rb is untouched, so selecting :data_dictionary retains exact master behavior.

Selection

OracleEnhancedAdapter.structure_dump_method accepts:

• :auto (default) — DBMS_METADATA on Oracle 12.1+, otherwise data-dictionary. The 12.1 floor is a project-policy version gate (the IDENTITY / EDITIONABLE / modern SET_TRANSFORM_PARAM era), exposed as use_dbms_metadata_dump?. DBMS_METADATA.GET_DDL itself works on Oracle 9i+; this is a project decision, not a strict capability check.
• :dbms_metadata — force DBMS_METADATA. Raises ArgumentError on pre-12.1 (mirrors PR Replace use_legacy_identifier_length with identifier_max_length (:auto default) #2576's identifier_max_length: :long fail-fast policy).
• :data_dictionary — force the implementation that assembles DDL from the ALL_* static data dictionary views in Ruby. Remains explicitly selectable on 12.1+ as well.

Implementation shape

The DBMS_METADATA path walks ALL_OBJECTS per object_type rather than hand-rolling per-kind queries. This naturally picks up MATERIALIZED VIEW, TYPE, TRIGGER, SYNONYM and friends without separate code paths and minimizes future maintenance when Oracle adds new object kinds. FK constraints (REF_CONSTRAINT) are emitted in a single trailing block; cross-table dependency order is left to the user (or to a fresh-schema reload) rather than tracked here.

Two modules in the same file split responsibilities so the dispatcher logic doesn't bleed into the DBMS_METADATA implementation:

• Dispatcher — prepend-ed; routes the three entry points (structure_dump, structure_dump_db_stored_code, structure_dump_synonyms) to either DBMS_METADATA or super (data-dictionary).
• DbmsMetadata — include-d; the DBMS_METADATA implementation itself, with no dispatch logic.

Output

Output shape is informed by mysqldump --no-data / pg_dump --schema-only:

• STORAGE / TABLESPACE / SEGMENT_ATTRIBUTES / EMIT_SCHEMA suppressed via DBMS_METADATA.SET_TRANSFORM_PARAM so the dump is portable across installations and schemas. Oracle honors these for TABLE / INDEX / SEQUENCE.
• MATERIALIZED VIEW is an Oracle-side exception: even with object_type-specific overrides Oracle does not honour the suppression on MV, so MV DDL retains physical attributes (PCTFREE, TABLESPACE, SEGMENT CREATION, etc.) as Oracle emits them. Since MVs aren't creatable through Rails' standard migration helpers anyway, no post-processing pass is added — Oracle's output is used as-is.
• REF_CONSTRAINTS emitted as separate ALTER TABLE … ADD CONSTRAINT statements after all tables.
• Inline constraints via Oracle's CONSTRAINTS=TRUE default; UNIQUE backing indexes that DBMS_METADATA already inlines into the table DDL are filtered out so they aren't emitted twice.
• STATEMENT_TOKEN separation preserved (Oracle's SQLTERMINATOR=FALSE default) so execute_structure_dump splits as before.
• primary_key_trigger: (Restore trigger-based primary key generation as opt-in primary_key_trigger: option #2615) row triggers are picked up via the TRIGGER object_type pass — addresses #2513 (comment).
• COMMENT ON TABLE / COMMENT ON COLUMN queried directly because GET_DEPENDENT_DDL("COMMENT", ...) is unreliable.
• structure_dump_db_stored_code selects 'PROCEDURE', 'PACKAGE', 'FUNCTION', 'TYPE' only — GET_DDL("PACKAGE", ...) already returns spec + body, so selecting 'PACKAGE BODY' would emit the body twice.
• Only SET_TRANSFORM_PARAM calls whose value differs from Oracle's default are issued (STORAGE / TABLESPACE / SEGMENT_ATTRIBUTES / EMIT_SCHEMA / REF_CONSTRAINTS).

Tests

A new dbms_metadata_structure_dump_spec.rb exercises the new path with structural assertions only (CREATE TABLE / CREATE INDEX / ALTER TABLE … ADD CONSTRAINT / COMMENT ON … / no STORAGE / TABLESPACE / PCTFREE noise) — exact byte-level DDL is intentionally not part of the contract, mirroring how pg_dump / mysqldump are tested upstream. Regression coverage added for the unique-index and PACKAGE-BODY duplicate-output bugs found during development. A schema-option spec verifies that :schema connection-time switching is honored (the dump walks SYS_CONTEXT('userenv', 'current_schema'), not the connecting user).

The existing structure_dump_spec.rb pins structure_dump_method = :data_dictionary in before(:all) and restores it in after(:all), so the data-dictionary backend's exact-DDL assertions continue to run.

Test plan

• bundle exec rspec spec/active_record/connection_adapters/oracle_enhanced/structure_dump_spec.rb — green (pinned to :data_dictionary)
• bundle exec rspec spec/active_record/connection_adapters/oracle_enhanced/dbms_metadata_structure_dump_spec.rb — green
• bundle exec rspec full suite — 747 examples, 0 failures, 12 pending
• bundle exec rubocop — clean

Closes #2513.

[yahonda]

Heads-up: #2615 (currently open) restores trigger-based primary key generation as the primary_key_trigger: opt-in. If #2615 lands, db:structure:dump should also pick up the BEFORE INSERT row trigger emitted by create_pk_trigger so that db:structure:load can recreate the table + sequence + trigger faithfully.

DBMS_METADATA.GET_DEPENDENT_DDL('TRIGGER', table_name, schema) returns the trigger DDL as a CLOB, which fits the GET_DDL-based approach this PR is taking. Worth adding when both PRs settle, either to this PR (if it stays open until #2615 merges) or as a follow-up.

The Ruby-flavored db/schema.rb round-trip is already handled inside #2615 via SchemaDumper; this note is purely about the db/structure.sql path.