diff --git a/sqlit/domains/connections/providers/adapters/base.py b/sqlit/domains/connections/providers/adapters/base.py
index 696c089..109ea1d 100644
--- a/sqlit/domains/connections/providers/adapters/base.py
+++ b/sqlit/domains/connections/providers/adapters/base.py
@@ -417,6 +417,22 @@ def quote_identifier(self, name: str) -> str:
         """Quote an identifier (table name, column name, etc.)."""
         pass
 
+    def qualified_name(self, database: str | None, schema: str | None, name: str) -> str:
+        """Build a quoted qualified identifier, skipping empty segments.
+
+        Default handles SQL Server-style `[db].[schema].[name]`, PostgreSQL-
+        style `"schema"."name"`, and single-part `"name"` by omitting any
+        empty/None component. Dialects that want different composition
+        (e.g. MySQL, which has no schemas within databases) can override.
+        """
+        parts: list[str] = []
+        if database:
+            parts.append(self.quote_identifier(database))
+        if schema:
+            parts.append(self.quote_identifier(schema))
+        parts.append(self.quote_identifier(name))
+        return ".".join(parts)
+
     @abstractmethod
     def build_select_query(self, table: str, limit: int, database: str | None = None, schema: str | None = None) -> str:
         """Build a SELECT query with limit.
diff --git a/sqlit/domains/connections/providers/model.py b/sqlit/domains/connections/providers/model.py
index 93b60b0..b104db3 100644
--- a/sqlit/domains/connections/providers/model.py
+++ b/sqlit/domains/connections/providers/model.py
@@ -62,6 +62,18 @@ def build_select_query(self, table: str, limit: int, database: str | None = None
 
     def format_table_name(self, schema: str | None, table: str) -> str: ...
 
+    def qualified_name(self, database: str | None, schema: str | None, name: str) -> str:
+        """Build a quoted qualified identifier for use in SQL.
+
+        Called when completing tables/views across multiple databases. Each
+        dialect decides how many segments to emit:
+          - SQL Server / generic: `[db].[schema].[name]`
+          - PostgreSQL:           `"schema"."name"` (databases are isolated)
+          - MySQL/MariaDB:        `` `db`.`name` `` (no schemas within databases)
+          - SQLite:               `"name"`
+        """
+        ...
+
 
 @runtime_checkable
 class SchemaInspector(Protocol):
diff --git a/sqlit/domains/query/ui/mixins/autocomplete_schema.py b/sqlit/domains/query/ui/mixins/autocomplete_schema.py
index a0e547d..bc8b7f4 100644
--- a/sqlit/domains/query/ui/mixins/autocomplete_schema.py
+++ b/sqlit/domains/query/ui/mixins/autocomplete_schema.py
@@ -406,13 +406,7 @@ def work() -> None:
                     if single_db:
                         full_name = table_name
                     else:
-                        quoted_db = dialect.quote_identifier(database) if database else ""
-                        quoted_schema = dialect.quote_identifier(schema_name)
-                        quoted_table = dialect.quote_identifier(table_name)
-                        if database:
-                            full_name = f"{quoted_db}.{quoted_schema}.{quoted_table}"
-                        else:
-                            full_name = f"{quoted_schema}.{quoted_table}"
+                        full_name = dialect.qualified_name(database, schema_name, table_name)
 
                     display_name = dialect.format_table_name(schema_name, table_name)
                     metadata = [
@@ -495,13 +489,7 @@ def work() -> None:
                     if single_db:
                         full_name = view_name
                     else:
-                        quoted_db = dialect.quote_identifier(database) if database else ""
-                        quoted_schema = dialect.quote_identifier(schema_name)
-                        quoted_view = dialect.quote_identifier(view_name)
-                        if database:
-                            full_name = f"{quoted_db}.{quoted_schema}.{quoted_view}"
-                        else:
-                            full_name = f"{quoted_schema}.{quoted_view}"
+                        full_name = dialect.qualified_name(database, schema_name, view_name)
 
                     display_name = dialect.format_table_name(schema_name, view_name)
                     metadata = [
@@ -816,14 +804,8 @@ async def run_db_call(fn: Any, *args: Any, **kwargs: Any) -> Any:
                             # Single database - use simple table name
                             schema_cache["tables"].append(table_name)
                         else:
-                            # Multiple databases - use full qualifier [db].[schema].[table]
-                            quoted_db = dialect.quote_identifier(database) if database else ""
-                            quoted_schema = dialect.quote_identifier(schema_name)
-                            quoted_table = dialect.quote_identifier(table_name)
-                            if database:
-                                full_name = f"{quoted_db}.{quoted_schema}.{quoted_table}"
-                            else:
-                                full_name = f"{quoted_schema}.{quoted_table}"
+                            # Multiple databases - use qualified identifier
+                            full_name = dialect.qualified_name(database, schema_name, table_name)
                             schema_cache["tables"].append(full_name)
                         # Keep metadata for column loading (multiple keys for flexible lookup)
                         display_name = dialect.format_table_name(schema_name, table_name)
@@ -843,14 +825,8 @@ async def run_db_call(fn: Any, *args: Any, **kwargs: Any) -> Any:
                             # Single database - use simple view name
                             schema_cache["views"].append(view_name)
                         else:
-                            # Multiple databases - use full qualifier [db].[schema].[view]
-                            quoted_db = dialect.quote_identifier(database) if database else ""
-                            quoted_schema = dialect.quote_identifier(schema_name)
-                            quoted_view = dialect.quote_identifier(view_name)
-                            if database:
-                                full_name = f"{quoted_db}.{quoted_schema}.{quoted_view}"
-                            else:
-                                full_name = f"{quoted_schema}.{quoted_view}"
+                            # Multiple databases - use qualified identifier
+                            full_name = dialect.qualified_name(database, schema_name, view_name)
                             schema_cache["views"].append(full_name)
                         # Keep metadata for column loading (multiple keys for flexible lookup)
                         display_name = dialect.format_table_name(schema_name, view_name)
diff --git a/sqlit/domains/query/ui/mixins/autocomplete_suggestions.py b/sqlit/domains/query/ui/mixins/autocomplete_suggestions.py
index 1cb1d58..94ef4e7 100644
--- a/sqlit/domains/query/ui/mixins/autocomplete_suggestions.py
+++ b/sqlit/domains/query/ui/mixins/autocomplete_suggestions.py
@@ -74,13 +74,25 @@ def _get_autocomplete_suggestions(self: AutocompleteMixinHost, text: str, cursor
             alias_map = self._build_alias_map(text)
             table_refs = extract_table_refs(text)
             loading: set[str] = getattr(self, "_columns_loading", set())
+            table_metadata = getattr(self, "_table_metadata", {}) or {}
+
+            def needs_column_load(key: str) -> bool:
+                """Return True only if the key names a known table that hasn't
+                been loaded yet. Returning Loading... for an unknown key would
+                wedge forever because the loader skips unknown tables silently.
+                """
+                if key in columns:
+                    return False
+                if key not in table_metadata:
+                    return False
+                return True
 
             for suggestion in suggestions:
                 if suggestion.type == SuggestionType.COLUMN:
                     # Check if any tables need column loading
                     for ref in table_refs:
                         table_key = ref.name.lower()
-                        if table_key not in columns and table_key not in loading:
+                        if needs_column_load(table_key) and table_key not in loading:
                             self._load_columns_for_table(table_key)
                             return ["Loading..."]
                         elif table_key in loading:
@@ -92,7 +104,7 @@ def _get_autocomplete_suggestions(self: AutocompleteMixinHost, text: str, cursor
                         scope_lower = scope.lower()
                         table_key = alias_map.get(scope_lower, scope_lower)
 
-                        if table_key not in columns and table_key not in loading:
+                        if needs_column_load(table_key) and table_key not in loading:
                             self._load_columns_for_table(table_key)
                             return ["Loading..."]
                         elif table_key in loading:
diff --git a/tests/unit/test_autocomplete_multidb.py b/tests/unit/test_autocomplete_multidb.py
new file mode 100644
index 0000000..e34946c
--- /dev/null
+++ b/tests/unit/test_autocomplete_multidb.py
@@ -0,0 +1,155 @@
+"""Regression tests for MySQL autocomplete in multi-database scenarios (#151).
+
+Two narrow bugs are covered:
+
+1. Qualified identifiers for databases without schemas (MySQL/MariaDB) used
+   to render as `db`.``.`table` — an empty-backticked middle segment. The
+   qualifying logic is now a Dialect method so each adapter owns its own
+   composition rule.
+
+2. Autocomplete returned a permanent "Loading..." sentinel whenever the
+   table reference in the query didn't resolve to anything in the schema
+   cache (e.g. the user typed `SELECT * FROM shop.cu`: `shop` was treated
+   as an alias and the loader spun forever for an unknown key).
+"""
+
+from __future__ import annotations
+
+import pytest
+
+
+# --------------------------------------------------------------------------
+# Bug 1: Dialect.qualified_name
+# --------------------------------------------------------------------------
+
+
+def _get_dialect(db_type: str):
+    from sqlit.domains.connections.providers.catalog import get_provider
+
+    return get_provider(db_type).dialect
+
+
+def test_mysql_qualified_name_is_two_part() -> None:
+    """MySQL has no schema-within-database; the qualified form is
+    `db`.`table`, NOT `db`.``.`table`."""
+    dialect = _get_dialect("mysql")
+    assert dialect.qualified_name("shop", "", "customers") == "`shop`.`customers`"
+
+
+def test_mariadb_qualified_name_is_two_part() -> None:
+    dialect = _get_dialect("mariadb")
+    assert dialect.qualified_name("shop", "", "customers") == "`shop`.`customers`"
+
+
+def test_postgresql_qualified_name_uses_schema_only() -> None:
+    """PostgreSQL databases are isolated; only schema.table makes sense
+    for cross-reference within the connected database."""
+    dialect = _get_dialect("postgresql")
+    # No db segment expected when schema is present.
+    assert dialect.qualified_name(None, "public", "users") == '"public"."users"'
+
+
+def test_sqlserver_qualified_name_is_three_part() -> None:
+    """SQL Server explicitly uses [db].[schema].[table]."""
+    dialect = _get_dialect("mssql")
+    assert dialect.qualified_name("app", "dbo", "Users") == "[app].[dbo].[Users]"
+
+
+def test_qualified_name_skips_empty_segments_everywhere() -> None:
+    """Regardless of dialect, empty db/schema segments must be omitted,
+    never rendered as empty-quoted placeholders."""
+    for db_type in ("mysql", "postgresql", "mssql", "sqlite"):
+        dialect = _get_dialect(db_type)
+        # bare table name: single quoted segment, no dot joins.
+        bare = dialect.qualified_name(None, None, "t")
+        assert "t" in bare, f"{db_type}: {bare}"
+        assert "." not in bare, f"{db_type} unexpected joins: {bare}"
+        # empty schema segment must not produce empty quote pair like `` or "" or [].
+        out = dialect.qualified_name(None, "", "t")
+        for marker in ("``", '""', "[]"):
+            assert marker not in out, f"{db_type} emitted empty-quoted segment: {out}"
+
+
+def test_qualified_name_escapes_embedded_quote_chars() -> None:
+    """Each dialect must escape its own quote char, not leak raw input."""
+    for db_type, payload, expected_substr in [
+        ("mysql", "app`evil", "app``evil"),
+        ("postgresql", 'app"evil', 'app""evil'),
+        ("mssql", "app]evil", "app]]evil"),
+    ]:
+        dialect = _get_dialect(db_type)
+        out = dialect.qualified_name(None, None, payload)
+        assert expected_substr in out, f"{db_type}: {out}"
+
+
+# --------------------------------------------------------------------------
+# Bug 2: stuck Loading... for unknown table references
+# --------------------------------------------------------------------------
+
+
+class _SchemaHost:
+    """Minimal stand-in for AutocompleteMixinHost — just enough to drive
+    `_get_autocomplete_suggestions`."""
+
+    def __init__(self, tables, metadata, columns=None, loading=None):
+        self._schema_cache = {
+            "tables": tables,
+            "views": [],
+            "columns": columns or {},
+            "procedures": [],
+        }
+        self._table_metadata = metadata
+        self._columns_loading = loading or set()
+        self.load_calls: list[str] = []
+
+    def _load_columns_for_table(self, table_name: str) -> None:
+        self.load_calls.append(table_name)
+
+    def _build_alias_map(self, text: str) -> dict:
+        return {}
+
+
+def test_unknown_table_ref_does_not_stick_on_loading() -> None:
+    """`SELECT * FROM shop.cu` parses `shop` as an ALIAS_COLUMN scope. It
+    isn't a real table. Before the fix the completion engine called
+    _load_columns_for_table('shop') and returned `Loading...`; the loader
+    skipped unknown keys silently, so the sentinel never cleared."""
+    from sqlit.domains.query.ui.mixins.autocomplete_suggestions import AutocompleteSuggestionsMixin
+
+    host = _SchemaHost(
+        tables=["customers", "orders", "products"],
+        metadata={
+            "customers": ("", "customers", "shop"),
+            "shop.customers": ("", "customers", "shop"),
+            "orders": ("", "orders", "shop"),
+            "products": ("", "products", "shop"),
+        },
+    )
+
+    get_suggestions = AutocompleteSuggestionsMixin._get_autocomplete_suggestions.__get__(host)
+
+    text = "SELECT * FROM shop.cu"
+    result = get_suggestions(text, len(text))
+
+    assert result != ["Loading..."], (
+        "unknown table key must not pin Loading... — loader never clears it"
+    )
+    assert "shop" not in host.load_calls
+
+
+def test_known_table_ref_still_triggers_loading_on_first_call() -> None:
+    """Sanity check: the fix must not regress legit lazy loading."""
+    from sqlit.domains.query.ui.mixins.autocomplete_suggestions import AutocompleteSuggestionsMixin
+
+    host = _SchemaHost(
+        tables=["customers"],
+        metadata={"customers": ("", "customers", None)},
+        columns={},
+    )
+    get_suggestions = AutocompleteSuggestionsMixin._get_autocomplete_suggestions.__get__(host)
+
+    text = "SELECT * FROM customers WHERE em"
+    result = get_suggestions(text, len(text))
+
+    assert result == ["Loading..."]
+    assert "customers" in host.load_calls