[osipovartem]

Why

SQL namespaces were represented only by property rows. Duplicate CREATE with disjoint property keys could succeed, empty namespaces required a synthetic property, and concurrent updates/drops could lose identity. SQL values were also interpolated into queries.

Change

• Add a dedicated iceberg_namespaces primary-key row. CREATE writes identity and properties atomically and reports unique violations as CommitConflict.
• Preserve property-only legacy namespaces. A small identity flag retains the old hidden exists marker without deleting user data; explicitly updated exists becomes visible.
• Bind namespace, table, property, and metadata-location values in SQLx queries for SQLite, PostgreSQL, and MySQL.
• Make namespace update/drop transactional. Lock the existing row or legacy property, avoid MySQL missing-row gap locks during promotion, and retry the whole transaction up to three times only for aborted deadlock/serialization SQLSTATEs.

Verification

• Local SQLite catalog tests: 20/20 passed; Rustice draft #954 executor create_schema 17/17 and catalog if_not_exists_ 2/2 passed against the equivalent runtime code.
• CI on the final head: Rust run 40266933328 passed lint and test. The SQL catalog job passed 23/23, including Docker-backed PostgreSQL and MySQL namespace-locking races and the existing full table integration test.
• Local Rust 1.95 targeted Clippy with -D warnings, cargo fmt --all --check, and git diff --check passed.
• Independent read-only code review approved the change, including a separate review of the final test-only lint fix.

Concurrent old-version writers should be quiesced during rollout. This migration reads old property-only catalogs but is not a mixed-version write protocol.