release: 0.12.0 - #26

Prepare 0.12.0 with fixes for query membership and NULLs, cursor and iterator boundaries, sliced prefetches, annotation names, bulk returning, tenant isolation, migration serialization and contrib model declarations. Add UUID version constraints, DTO extra/related fields and migration safety checks. Update the English and Russian documentation and consolidate regression coverage into contract matrices.

SQLite dictionary reads preserve case-distinct output names. Values set operations use internal positional column names when wrapping branches, preserving output names through slicing, nesting and cursor streaming.

Recursive queries declared before initialization replay their calls after model setup, so unsupported combinations are reported by init with their declaration location. Tests initialize their models and lazy dialect registries explicitly before checking errors or plan reuse, removing dependence on test order without weakening assertions.

Validation:

• Full SQLite pure-Python suite: 13,171 passed, 2,154 skipped, with plan verification.
• Recursive/deferred query and plan tests on Python 3.12: 549 passed on Windows and Linux; PostgreSQL/asyncpg: 571 passed.
• Earlier checks of the case-distinct output fix: PostgreSQL 109 passed; ClickHouse 40 passed.
• make _check passed on Windows. Linux formatting, lint, mypy, Bandit, focused tests and strict documentation build passed.
• The 18-operation PostgreSQL rust comparison retained Python calls and SQL counts. Recursive reads retain 280 Python calls and one SQL per operation; repeated paired timings show no consistent slowdown.

The full PR CI matrix must validate this updated commit before merge.