refactor(persistence): freeze legacy DB migration chain, drop dbm026

The legacy pkg/persistence/migrations (DBMigration / dbmXXX) system now
coexists with Alembic but accepts no new migrations — all new schema
changes go through Alembic.

- remove dbm026_llm_model_context_length (superseded by Alembic
  0005_add_llm_context_length, which makes the identical change)
- cap required_database_version at 25 (legacy chain dbm001-025 kept
  read-only to upgrade pre-existing 3.x DBs to the Alembic baseline)
- add migrations/README.md documenting the freeze
- document the Alembic-only policy and revision-id/idempotency rules in
  AGENTS.md

src/langbot/pkg/persistence/migrations/README.md

@@ -0,0 +1,36 @@
+# Legacy migrations (DEPRECATED — do not add new files here)
+
+This directory holds the **legacy 3.x database migration system**
+(`DBMigration` subclasses in `dbmXXX_*.py`, registered via
+`@migration.migration_class(N)` and run from `pkg/persistence/mgr.py`).
+
+**This system is frozen. Do not add new `dbmXXX_*.py` migrations.**
+
+The chain is capped at version 25 (`required_database_version = 25` in
+`pkg/utils/constants.py`). These files exist only to upgrade pre-existing
+3.x databases up to the Alembic baseline (`0001_baseline`). Removing them
+would break in-place upgrades from old installations, so they are kept
+read-only.
+
+## All new schema changes use Alembic
+
+Migrations now live in `pkg/persistence/alembic/versions/`. To create one:
+
+```bash
+uv run python -m langbot.pkg.persistence.alembic_runner autogenerate "description of your change"
+```
+
+(requires `data/config.yaml` to exist). Review and edit the generated
+script before committing — Alembic migrations run automatically on startup
+and must be idempotent and guard against missing tables (the test suite
+runs them against empty databases).
+
+### Rules for Alembic revision ids
+
+- Keep the revision id **≤ 32 characters** — PostgreSQL stores
+  `alembic_version.version_num` as `varchar(32)` and will raise
+  `StringDataRightTruncationError` on overflow.
+- Guard every `op` call against a missing table / missing column
+  (`inspector.get_table_names()` / `inspector.get_columns()`); fresh
+  installs create the schema via `create_all()` and stamp the baseline,
+  so migrations may run against tables that already match or do not exist.

src/langbot/pkg/persistence/migrations/dbm026_llm_model_context_length.py

@@ -1,42 +0,0 @@
-import sqlalchemy
-from .. import migration
-
-
-@migration.migration_class(26)
-class DBMigrateLLMModelContextLength(migration.DBMigration):
-    """Add context_length column to LLM models"""
-
-    async def upgrade(self):
-        columns = await self._get_columns('llm_models')
-        if 'context_length' not in columns:
-            await self.ap.persistence_mgr.execute_async(
-                sqlalchemy.text('ALTER TABLE llm_models ADD COLUMN context_length INTEGER')
-            )
-
-    async def downgrade(self):
-        columns = await self._get_columns('llm_models')
-        if 'context_length' not in columns:
-            return
-
-        if self.ap.persistence_mgr.db.name == 'postgresql':
-            await self.ap.persistence_mgr.execute_async(
-                sqlalchemy.text('ALTER TABLE llm_models DROP COLUMN IF EXISTS context_length')
-            )
-        else:
-            await self.ap.persistence_mgr.execute_async(
-                sqlalchemy.text('ALTER TABLE llm_models DROP COLUMN context_length')
-            )
-
-    async def _get_columns(self, table_name: str) -> set[str]:
-        if self.ap.persistence_mgr.db.name == 'postgresql':
-            result = await self.ap.persistence_mgr.execute_async(
-                sqlalchemy.text("""
-                    SELECT column_name FROM information_schema.columns
-                    WHERE table_name = :table_name
-                """),
-                {'table_name': table_name},
-            )
-            return {row[0] for row in result.fetchall()}
-
-        result = await self.ap.persistence_mgr.execute_async(sqlalchemy.text(f'PRAGMA table_info({table_name})'))
-        return {row[1] for row in result.fetchall()}