feat(core): Improve escaping in destinations and rework state structure

[iambriccardo]

Summary

This PR improves destination SQL escaping/quoting, makes replica-identity failures easier to classify and remediate, and cleans up table state naming across core ETL and the Postgres-backed persistence layer.

Changes

Destination SQL quoting and escaping

• Adds destination-specific SQL helper modules:
  • bigquery/sql.rs for GoogleSQL identifier/path quoting.
  • clickhouse/sql.rs for ClickHouse identifier quoting.
  • ducklake/sql.rs for DuckDB/DuckLake identifier quoting and lake catalog qualification.
  • snowflake/sql.rs for Snowflake identifiers and string literals.
  • sql.rs for shared ANSI double-quoted identifier escaping.
• Moves BigQuery identifier escaping out of BigQueryClient and reuses it for fully qualified table paths, INFORMATION_SCHEMA.TABLES paths, column DDL, and primary key clauses.
• Updates ClickHouse DDL/test utilities to use the new ClickHouse quoting helper, including create/drop database statements.
• Updates DuckLake SQL generation to quote and qualify target table names consistently instead of string-replacing generated DDL.
• Updates DuckLake mutation SQL to quote column names in delete predicates, update assignments, staging table operations, truncate, delete, update, and insert-select paths.
• Keeps Postgres metadata SQL in DuckLake explicit by renaming that helper to quote_postgres_identifier, separating Postgres metadata SQL from DuckDB/DuckLake SQL.
• Updates Snowflake table existence checks to use SHOW TERSE TABLES ... STARTS WITH ... plus exact row matching, with centralized string literal quoting.

Replica identity error classification

• Adds ErrorKind::SourceReplicaIdentityError.
• Reclassifies destination replica-identity incompatibilities from broader error kinds to the new source replica identity error kind across BigQuery, ClickHouse, DuckLake, Iceberg, and Snowflake.
• Adds a specific manual remediation policy for replica-identity errors explaining how to choose the least costly supported replica identity:
  • REPLICA IDENTITY DEFAULT with a primary key.
  • REPLICA IDENTITY USING INDEX when supported.
  • REPLICA IDENTITY FULL only when full row images are needed.
• Adds a BigQuery helper for partial update row errors so those failures also use the replica identity error kind.

Table state structure and naming

• Flattens the core etl::state module so table-specific state lives directly under state/:
  • state/destination_table_metadata.rs.
  • state/table_error.rs.
  • state/table_retry_policy.rs.
  • state/table_state.rs.
• Renames the core ETL domain types to state-oriented names:
  • TableState.
  • TableStateType.
  • TableError.
  • TableRetryPolicy.
• Flattens the public state API so callers use etl::state::{TableState, TableStateType, TableError, TableRetryPolicy} instead of a nested table module.
• Renames StateStore table APIs to state-oriented names:
  • get_table_state.
  • get_table_states.
  • load_table_states.
  • update_table_state.
  • update_table_states.
  • rollback_table_state.
• Renames the cached table map alias from TableReplicationStates to TableStates.
• Aligns ETL Postgres persistence modules with the same concepts:
  • destination_metadata -> destination_table_metadata.
  • state -> table_state.
• Keeps the persisted Postgres representation clearly separate from the core domain model:
  • TableStateRow represents the database row.
  • StoredTableStateType represents the database enum value.
  • Core ETL code converts with TableState::to_storage_format() and TableState::from_state_row(...).
• Keeps database-facing functions row/state-oriented (get_table_state_rows, update_table_state_raw, rollback_table_state, reset_table_state, delete_table_state(s)) while exposing core ETL state values to the rest of the pipeline.
• Simplifies TableError so it stores replication failure details and the source error without carrying a table id.
• Reuses the shared worker error handling policy when building table errors.
• Adds Hash support and conversions for TableStateType.
• Converts table state metrics from string-keyed counts to TableStateType-keyed counts, converting to the static metric label only at emission time.
• Renames the table-state metric label from phase to state.
• Updates table-state logs, comments, and public docs so they no longer refer to table replication phases.
• Renames API schema/internal symbols from the old table-replication-state vocabulary to SimpleTableState, TableStatus, MissingTableState, and InvalidTableState.
• Updates API rollback/status and replicator error reporting call sites to bridge from persisted TableStateRow metadata into core TableState values.

Migration and persistence notes

• Leaves migration files unchanged relative to origin/main; the table-state rename is code-level and does not require a database migration.
• Keeps the persisted table-state JSON representation on the same serde-tagged shape (type, lsn, reason, solution, and retry_policy).
• Keeps source_err memory-only and skipped from persisted status JSON, so newly added or removed internal ErrorKind variants are not stored in table state.
• Adds an AGENTS.md rule that migration files, including comments and non-executable text, must not be modified unless explicitly requested.

[coveralls]

coverage: 73.138% (+0.2%) from 72.926% — improve-code into main

[farazdagi]

Snowflake destination definitely looks nicer with that sql.rs extracted.