MCPg v0.8.2

Added

• mcpg.identifiers — one escape-based SQL identifier quoter (quote_identifier / ensure_identifier).
  It doubles embedded double quotes (" → "") so any name PostgreSQL accepts as a delimited
  identifier can be spliced safely, and rejects only the values that are not addressable identifiers at
  all: the empty string, an embedded NUL, and names over PostgreSQL's 63-byte limit (which the server
  would silently truncate). It replaces the copy-pasted [A-Za-z_][A-Za-z0-9_]* regex that ~20 modules
  each carried, where that regex was doing double duty as both the injection guard and (accidentally)
  a blanket ban on every hyphen / space / mixed-case / quoted name. Covered by an adversarial test suite
  (tests/unit/test_identifiers.py).

Fixed

• Tools rejected valid PostgreSQL object names that need delimited-identifier quoting
  (#329). The reported case was dump_database: a real
  schema such as adm-pgbench failed with ShellError: invalid schema name: 'adm-pgbench' before
  pg_dump was ever spawned. pg_dump --schema takes a pattern (psql \d rules) — an unquoted
  * / ? / [ is a wildcard and bare letters fold to lowercase — so each schemas entry is now
  encoded as a double-quoted literal pattern (_encode_schema_pattern); copy_table_between_databases
  encodes its --table the same way. The access-mode requirement (unrestricted + MCPG_ALLOW_SHELL)
  and the DumpResult contract are unchanged.

  A sweep fixed the same class of over-restriction across the in-process SQL tools, which now quote
  via mcpg.identifiers instead of validating against the plain-identifier allowlist:
  export_table / import_csv / import_json / import_vectors (data movement), the pgvector suites
  (vector_ops, vector_tuning, rag_efficiency, pg_search, turboquant), textsearch,
  composite, rls and tenancy / config roles (SET LOCAL ROLE), logical_replication,
  migrations, test_data / test_row_factory, redis_fdw, listen (LISTEN/NOTIFY channels), and
  timescaledb (whose create_hypertable / policy calls take the relation as a regclass text
  argument, so the quoted relation is wrapped in a single-quoted literal with both layers escaped).
  Hyphens, spaces, mixed case, and embedded quotes now work and are safely escaped everywhere; only
  empty / NUL / overlong names are rejected.

  Left deliberately strict, because the name is not being used as a PostgreSQL identifier there:
  the Apache AGE graph tools (graph, graph_projection, cypher, graph_diagram), where labels
  follow AGE's own naming rules; the ORM/query-builder code generators (prisma, drizzle, diesel,
  ecto, ent, jooq, sqlc, sqlalchemy_export), where the name becomes an identifier in generated
  Go/Rust/Elixir/Python source; and SQL/PGQ property-graph names (pgq). textsearch also keeps the
  strict allowlist for the one value it embeds in a string literal (a regconfig name).