feat: add Alembic for database migrations

Replace inline migration logic with proper Alembic migrations:
- 001: Initial schema (creates memories table with FTS)
- 002: Add multi-user and secrets columns (user_id, is_sensitive,
  vault_path, encrypted_content)

Migrations run automatically on app startup. Existing databases
are handled gracefully with IF NOT EXISTS / column existence checks.

migrations/versions/001_initial_schema.py

@@ -0,0 +1,50 @@
+"""Initial schema with memories table.
+
+Revision ID: 001
+Revises:
+Create Date: 2026-03-14
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = "001"
+down_revision: Union[str, None] = None
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    # Check if table already exists (handles pre-Alembic installations)
+    result = conn.execute(
+        sa.text("SELECT EXISTS(SELECT 1 FROM information_schema.tables WHERE table_name = 'memories')")
+    )
+    if result.scalar():
+        return
+
+    op.execute("""
+        CREATE TABLE memories (
+            id SERIAL PRIMARY KEY,
+            content TEXT NOT NULL,
+            category VARCHAR(50) DEFAULT 'facts',
+            tags TEXT DEFAULT '',
+            expanded_keywords TEXT DEFAULT '',
+            importance REAL DEFAULT 0.5,
+            created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
+            updated_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
+            search_vector tsvector GENERATED ALWAYS AS (
+                setweight(to_tsvector('english', coalesce(content, '')), 'A') ||
+                setweight(to_tsvector('english', coalesce(expanded_keywords, '')), 'B') ||
+                setweight(to_tsvector('english', coalesce(tags, '')), 'C') ||
+                setweight(to_tsvector('english', coalesce(category, '')), 'D')
+            ) STORED
+        )
+    """)
+    op.execute("CREATE INDEX idx_memories_search ON memories USING GIN(search_vector)")
+
+
+def downgrade() -> None:
+    op.drop_index("idx_memories_search")
+    op.drop_table("memories")