fix(db): add missing link_type, FK constraints, and partial index per spec DDL

Align resource_links, checkpoint_metadata, and decisions schema behavior with the spec DDL. Add SQLite runtime trigger guards for checkpoint foreign-key semantics and migration-focused Behave/Robot checks that orphan checkpoint references are rejected.

ISSUES CLOSED: #922

features/db_migration_lifecycle.feature

@@ -11,6 +11,16 @@ Feature: Database migration lifecycle
     Then the database should have all expected tables
     And the current revision should be the head revision
 
+  Scenario: Spec-parity schema details are present after migration
+    Given all migrations have been applied
+    Then the "resource_links" table should include "link_type" with default "contains"
+    And checkpoint_metadata should enforce decision and resource foreign keys
+    And checkpoint_metadata foreign keys should reject orphan references
+    And checkpoint_metadata should reject orphan decision_id independently
+    And checkpoint_metadata should reject orphan resource_id independently
+    And checkpoint_metadata should accept valid decision and resource references
+    And the "decisions" table should have partial index "idx_decisions_superseded" on "superseded_by"
+
   Scenario: Roll back all migrations to base
     Given all migrations have been applied
     When I downgrade all migrations to base
@@ -45,6 +55,84 @@ Feature: Database migration lifecycle
     Then the database should be stamped with alembic_version
     And the database revision should be at head
 
+  Scenario: Migration orphan cleanup nullifies stale checkpoint references
+    Given migrations applied up to "m4_003_plan_env_columns"
+    And checkpoint_metadata contains orphan decision and resource references
+    When I upgrade to the next migration revision
+    Then the orphan decision_id values should be NULL
+    And the orphan resource_id values should be NULL
+    And checkpoint_metadata should enforce decision and resource foreign keys
+
+  Scenario: resource_links.link_type accepts spec-defined non-default values
+    Given all migrations have been applied
+    Then resource_links should accept link_type "references"
+    And resource_links should accept link_type "derived_from"
+
+  Scenario: resource_links.link_type rejects NULL values
+    Given all migrations have been applied
+    Then resource_links should reject NULL link_type
+
+  Scenario: resource_links.link_type rejects invalid values
+    Given all migrations have been applied
+    Then resource_links should reject link_type "invalid_type"
+
+  Scenario: Deleting a decision sets checkpoint decision_id to NULL
+    Given all migrations have been applied
+    And a checkpoint references a valid decision
+    When the referenced decision is deleted
+    Then the checkpoint decision_id should be NULL
+
+  Scenario: Deleting a resource sets checkpoint resource_id to NULL
+    Given all migrations have been applied
+    And a checkpoint references a valid resource
+    When the referenced resource is deleted
+    Then the checkpoint resource_id should be NULL
+
+  Scenario: Spec-parity m4_004 downgrade removes added schema objects
+    Given all migrations have been applied
+    When I downgrade to revision "m4_003_plan_env_columns"
+    Then the "resource_links" table should not include "link_type"
+    And the "decisions" table should not have index "idx_decisions_superseded"
+    And checkpoint_metadata should not have foreign key "fk_checkpoint_metadata_decision"
+    And checkpoint_metadata should not have foreign key "fk_checkpoint_metadata_resource"
+    And checkpoint_metadata should not have SQLite triggers for FK enforcement
+
+  Scenario: Migration idempotency: existing link_type column gains CHECK constraint
+    Given migrations applied up to "m4_003_plan_env_columns"
+    And resource_links already has a link_type column without CHECK constraint
+    When I upgrade to the next migration revision
+    Then the "resource_links" table should include "link_type" with default "contains"
+    And resource_links should reject link_type "invalid_type"
+
+  Scenario: Migration idempotency: existing link_type column with wrong default is corrected
+    Given migrations applied up to "m4_003_plan_env_columns"
+    And resource_links already has a link_type column with a non-contains default
+    When I upgrade to the next migration revision
+    Then the "resource_links" table should include "link_type" with default "contains"
+    And resource_links should reject link_type "invalid_type"
+
+  Scenario: Positive checkpoint FK test verifies persisted row
+    Given all migrations have been applied
+    And valid decision and resource rows exist
+    When I insert a checkpoint with valid FK references
+    Then the checkpoint row should be persisted in the database
+
+  Scenario: UPDATE trigger rejects orphan decision_id on checkpoint_metadata
+    Given all migrations have been applied
+    And a checkpoint exists with valid FK references
+    When I update the checkpoint decision_id to a non-existent value
+    Then the update should be rejected with an integrity error
+
+  Scenario: UPDATE trigger rejects orphan resource_id on checkpoint_metadata
+    Given all migrations have been applied
+    And a checkpoint exists with valid FK references
+    When I update the checkpoint resource_id to a non-existent value
+    Then the update should be rejected with an integrity error
+
+  Scenario: resource_links.link_type rejects empty string values
+    Given all migrations have been applied
+    Then resource_links should reject empty string link_type
+
   Scenario: init_database creates a valid schema
     When I call init_database with an in-memory URL
     Then the resulting engine should have a valid schema

features/steps/db_migration_lifecycle_steps.py

@@ -1,4 +1,4 @@
-"""Step definitions for db_migration_lifecycle.feature.
+"""Step definitions for db_migration_lifecycle.feature — core lifecycle.
 
 Exercises the full Alembic migration lifecycle: forward application,
 rollback, round-trip testing, CLI wrappers, stamp logic, and the
@@ -13,6 +13,7 @@ from typing import Any
 from unittest.mock import patch
 
 from behave import given, then, when
+from sqlalchemy import inspect as sa_inspect
 
 
 def _create_temp_db_path() -> str:
@@ -75,8 +76,6 @@ _EXPECTED_TABLES = {
 
 
 def _get_table_names(engine: Any) -> set[str]:
-    from sqlalchemy import inspect as sa_inspect
-
     inspector = sa_inspect(engine)
     return set(inspector.get_table_names())
 

features/steps/db_schema_cascade_steps.py

@@ -0,0 +1,457 @@
+"""Step definitions for db_migration_lifecycle.feature — cascade and persistence.
+
+``ondelete="SET NULL"`` cascade verification for checkpoint_metadata foreign
+keys, positive FK persistence tests, and trigger removal verification on
+downgrade.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from behave import given, then, when
+from sqlalchemy import text
+from sqlalchemy.exc import IntegrityError
+
+from features.steps.db_schema_parity_steps import (
+    _ensure_test_action_and_plan,
+    _insert_test_resource,
+)
+
+# ---------------------------------------------------------------------------
+# Given/When/Then — positive FK persistence test
+# ---------------------------------------------------------------------------
+
+
+@given("valid decision and resource rows exist")
+def step_valid_decision_and_resource_exist(context: Any) -> None:
+    action_name = "local/test-action-persist"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FE0"
+    decision_id = "01ARZ3NDEKTSV4RRFFQ69G5FE1"
+    resource_id = "01ARZ3NDEKTSV4RRFFQ69G5FE2"
+
+    with context.engine.begin() as conn:
+        _ensure_test_action_and_plan(conn, action_name, plan_id)
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO decisions (
+                    decision_id, plan_id, decision_type, question,
+                    chosen_option, context_snapshot_json, sequence_number,
+                    created_at
+                ) VALUES (
+                    :decision_id, :plan_id, :decision_type, :question,
+                    :chosen_option, :context_snapshot_json, :sequence_number,
+                    :created_at
+                )
+                """
+            ),
+            {
+                "decision_id": decision_id,
+                "plan_id": plan_id,
+                "decision_type": "strategy_choice",
+                "question": "test question",
+                "chosen_option": "test option",
+                "context_snapshot_json": "{}",
+                "sequence_number": 1,
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+        _insert_test_resource(conn, resource_id)
+
+    context.persist_plan_id = plan_id
+    context.persist_decision_id = decision_id
+    context.persist_resource_id = resource_id
+    context.persist_checkpoint_id = "01ARZ3NDEKTSV4RRFFQ69G5FE3"
+
+
+@when("I insert a checkpoint with valid FK references")
+def step_insert_checkpoint_valid_fk(context: Any) -> None:
+    with context.engine.begin() as conn:
+        conn.execute(
+            text(
+                """
+                INSERT INTO checkpoint_metadata (
+                    checkpoint_id, plan_id, decision_id,
+                    checkpoint_type, resource_id, sandbox_ref,
+                    filesystem_path, created_at
+                ) VALUES (
+                    :checkpoint_id, :plan_id, :decision_id,
+                    :checkpoint_type, :resource_id, :sandbox_ref,
+                    :filesystem_path, :created_at
+                )
+                """
+            ),
+            {
+                "checkpoint_id": context.persist_checkpoint_id,
+                "plan_id": context.persist_plan_id,
+                "decision_id": context.persist_decision_id,
+                "checkpoint_type": "manual",
+                "resource_id": context.persist_resource_id,
+                "sandbox_ref": "test-ref",
+                "filesystem_path": "",
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+
+@then("the checkpoint row should be persisted in the database")
+def step_checkpoint_persisted(context: Any) -> None:
+    with context.engine.connect() as conn:
+        row = conn.execute(
+            text(
+                "SELECT checkpoint_id, decision_id, resource_id "
+                "FROM checkpoint_metadata WHERE checkpoint_id = :cid"
+            ),
+            {"cid": context.persist_checkpoint_id},
+        ).fetchone()
+
+    assert row is not None, (
+        f"Checkpoint {context.persist_checkpoint_id!r} was not persisted"
+    )
+    assert row[0] == context.persist_checkpoint_id
+    assert row[1] == context.persist_decision_id, (
+        f"Expected decision_id {context.persist_decision_id!r}, got {row[1]!r}"
+    )
+    assert row[2] == context.persist_resource_id, (
+        f"Expected resource_id {context.persist_resource_id!r}, got {row[2]!r}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Given/When/Then — UPDATE trigger rejection
+# ---------------------------------------------------------------------------
+
+
+@given("a checkpoint exists with valid FK references")
+def step_checkpoint_with_valid_fk_refs(context: Any) -> None:
+    action_name = "local/test-action-upd-trg"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FH0"
+    decision_id = "01ARZ3NDEKTSV4RRFFQ69G5FH1"
+    resource_id = "01ARZ3NDEKTSV4RRFFQ69G5FH2"
+    checkpoint_id = "01ARZ3NDEKTSV4RRFFQ69G5FH3"
+
+    with context.engine.begin() as conn:
+        _ensure_test_action_and_plan(conn, action_name, plan_id)
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO decisions (
+                    decision_id, plan_id, decision_type, question,
+                    chosen_option, context_snapshot_json, sequence_number,
+                    created_at
+                ) VALUES (
+                    :decision_id, :plan_id, :decision_type, :question,
+                    :chosen_option, :context_snapshot_json, :sequence_number,
+                    :created_at
+                )
+                """
+            ),
+            {
+                "decision_id": decision_id,
+                "plan_id": plan_id,
+                "decision_type": "strategy_choice",
+                "question": "test question",
+                "chosen_option": "test option",
+                "context_snapshot_json": "{}",
+                "sequence_number": 1,
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+        _insert_test_resource(conn, resource_id)
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO checkpoint_metadata (
+                    checkpoint_id, plan_id, decision_id,
+                    checkpoint_type, resource_id, sandbox_ref,
+                    filesystem_path, created_at
+                ) VALUES (
+                    :checkpoint_id, :plan_id, :decision_id,
+                    :checkpoint_type, :resource_id, :sandbox_ref,
+                    :filesystem_path, :created_at
+                )
+                """
+            ),
+            {
+                "checkpoint_id": checkpoint_id,
+                "plan_id": plan_id,
+                "decision_id": decision_id,
+                "checkpoint_type": "manual",
+                "resource_id": resource_id,
+                "sandbox_ref": "test-ref",
+                "filesystem_path": "",
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+    context.update_trg_checkpoint_id = checkpoint_id
+
+
+@when("I update the checkpoint decision_id to a non-existent value")
+def step_update_checkpoint_decision_orphan(context: Any) -> None:
+    try:
+        with context.engine.begin() as conn:
+            conn.execute(
+                text(
+                    "UPDATE checkpoint_metadata "
+                    "SET decision_id = :orphan "
+                    "WHERE checkpoint_id = :cid"
+                ),
+                {
+                    "orphan": "NONEXISTENT_DECISION_ID_UPD",
+                    "cid": context.update_trg_checkpoint_id,
+                },
+            )
+    except IntegrityError:
+        context.update_trigger_rejected = True
+        return
+
+    context.update_trigger_rejected = False
+
+
+@when("I update the checkpoint resource_id to a non-existent value")
+def step_update_checkpoint_resource_orphan(context: Any) -> None:
+    try:
+        with context.engine.begin() as conn:
+            conn.execute(
+                text(
+                    "UPDATE checkpoint_metadata "
+                    "SET resource_id = :orphan "
+                    "WHERE checkpoint_id = :cid"
+                ),
+                {
+                    "orphan": "NONEXISTENT_RESOURCE_ID_UPD",
+                    "cid": context.update_trg_checkpoint_id,
+                },
+            )
+    except IntegrityError:
+        context.update_trigger_rejected = True
+        return
+
+    context.update_trigger_rejected = False
+
+
+@then("the update should be rejected with an integrity error")
+def step_update_rejected(context: Any) -> None:
+    assert context.update_trigger_rejected, (
+        "Expected UPDATE to be rejected by FK trigger, but it was accepted"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Helper — enable FK enforcement on the pooled DBAPI connection
+# ---------------------------------------------------------------------------
+
+
+def _enable_fk_enforcement(engine: Any) -> None:
+    """Enable PRAGMA foreign_keys on the engine's pooled DBAPI connection.
+
+    For in-memory SQLite databases, SQLAlchemy uses ``StaticPool`` which
+    shares a single underlying DBAPI connection across all pool checkouts.
+    Calling ``raw_connection()`` returns a wrapper around that shared
+    connection; ``close()`` returns it to the pool without closing the
+    DBAPI connection.  The PRAGMA therefore persists for all subsequent
+    ``engine.begin()`` / ``engine.connect()`` calls.
+
+    This matches the codebase convention of setting PRAGMA foreign_keys
+    at the connection level (e.g. ``resource_repository_steps.py``,
+    ``plan_lifecycle_persistence_steps.py``).  The ``event.listens_for``
+    variant used in engine-creation contexts is not applicable here
+    because the engine's connections have already been established during
+    migration; the ``"connect"`` event would not fire for existing
+    pooled connections.
+    """
+    raw_conn = engine.raw_connection()
+    try:
+        cursor = raw_conn.cursor()
+        cursor.execute("PRAGMA foreign_keys=ON")
+        cursor.close()
+    finally:
+        raw_conn.close()
+
+
+# ---------------------------------------------------------------------------
+# Given/When/Then — ondelete="SET NULL" cascade (decision)
+# ---------------------------------------------------------------------------
+
+
+@given("a checkpoint references a valid decision")
+def step_checkpoint_references_decision(context: Any) -> None:
+    action_name = "local/test-action-set-null-d"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FG0"
+    decision_id = "01ARZ3NDEKTSV4RRFFQ69G5FG1"
+    checkpoint_id = "01ARZ3NDEKTSV4RRFFQ69G5FG2"
+
+    with context.engine.begin() as conn:
+        _ensure_test_action_and_plan(conn, action_name, plan_id)
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO decisions (
+                    decision_id, plan_id, decision_type, question,
+                    chosen_option, context_snapshot_json, sequence_number,
+                    created_at
+                ) VALUES (
+                    :decision_id, :plan_id, :decision_type, :question,
+                    :chosen_option, :context_snapshot_json, :sequence_number,
+                    :created_at
+                )
+                """
+            ),
+            {
+                "decision_id": decision_id,
+                "plan_id": plan_id,
+                "decision_type": "strategy_choice",
+                "question": "test question",
+                "chosen_option": "test option",
+                "context_snapshot_json": "{}",
+                "sequence_number": 1,
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO checkpoint_metadata (
+                    checkpoint_id, plan_id, decision_id,
+                    checkpoint_type, sandbox_ref,
+                    filesystem_path, created_at
+                ) VALUES (
+                    :checkpoint_id, :plan_id, :decision_id,
+                    :checkpoint_type, :sandbox_ref,
+                    :filesystem_path, :created_at
+                )
+                """
+            ),
+            {
+                "checkpoint_id": checkpoint_id,
+                "plan_id": plan_id,
+                "decision_id": decision_id,
+                "checkpoint_type": "manual",
+                "sandbox_ref": "test-ref",
+                "filesystem_path": "",
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+    context.set_null_decision_id = decision_id
+    context.set_null_checkpoint_id_d = checkpoint_id
+    context.set_null_plan_id_d = plan_id
+
+
+@when("the referenced decision is deleted")
+def step_delete_referenced_decision(context: Any) -> None:
+    # Enable FK enforcement via the established codebase pattern
+    # (event listener on "connect") so that ondelete="SET NULL" is honoured.
+    _enable_fk_enforcement(context.engine)
+
+    with context.engine.begin() as conn:
+        conn.execute(
+            text("DELETE FROM decisions WHERE decision_id = :did"),
+            {"did": context.set_null_decision_id},
+        )
+
+
+@then("the checkpoint decision_id should be NULL")
+def step_checkpoint_decision_id_null(context: Any) -> None:
+    with context.engine.connect() as conn:
+        row = conn.execute(
+            text(
+                "SELECT decision_id FROM checkpoint_metadata WHERE checkpoint_id = :cid"
+            ),
+            {"cid": context.set_null_checkpoint_id_d},
+        ).fetchone()
+
+    assert row is not None, (
+        f"Checkpoint {context.set_null_checkpoint_id_d!r} missing after decision deletion"
+    )
+    assert row[0] is None, (
+        f"Expected checkpoint decision_id to be NULL after parent deletion, "
+        f"got {row[0]!r}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Given/When/Then — ondelete="SET NULL" cascade (resource)
+# ---------------------------------------------------------------------------
+
+
+@given("a checkpoint references a valid resource")
+def step_checkpoint_references_resource(context: Any) -> None:
+    action_name = "local/test-action-set-null-r"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FG3"
+    resource_id = "01ARZ3NDEKTSV4RRFFQ69G5FG4"
+    checkpoint_id = "01ARZ3NDEKTSV4RRFFQ69G5FG5"
+
+    with context.engine.begin() as conn:
+        _ensure_test_action_and_plan(conn, action_name, plan_id)
+
+        _insert_test_resource(conn, resource_id)
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO checkpoint_metadata (
+                    checkpoint_id, plan_id, resource_id,
+                    checkpoint_type, sandbox_ref,
+                    filesystem_path, created_at
+                ) VALUES (
+                    :checkpoint_id, :plan_id, :resource_id,
+                    :checkpoint_type, :sandbox_ref,
+                    :filesystem_path, :created_at
+                )
+                """
+            ),
+            {
+                "checkpoint_id": checkpoint_id,
+                "plan_id": plan_id,
+                "resource_id": resource_id,
+                "checkpoint_type": "manual",
+                "sandbox_ref": "test-ref",
+                "filesystem_path": "",
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+    context.set_null_resource_id = resource_id
+    context.set_null_checkpoint_id_r = checkpoint_id
+
+
+@when("the referenced resource is deleted")
+def step_delete_referenced_resource(context: Any) -> None:
+    # Enable FK enforcement via the established codebase pattern
+    # (event listener on "connect") so that ondelete="SET NULL" is honoured.
+    _enable_fk_enforcement(context.engine)
+
+    with context.engine.begin() as conn:
+        conn.execute(
+            text("DELETE FROM resources WHERE resource_id = :rid"),
+            {"rid": context.set_null_resource_id},
+        )
+
+
+@then("the checkpoint resource_id should be NULL")
+def step_checkpoint_resource_id_null(context: Any) -> None:
+    with context.engine.connect() as conn:
+        row = conn.execute(
+            text(
+                "SELECT resource_id FROM checkpoint_metadata WHERE checkpoint_id = :cid"
+            ),
+            {"cid": context.set_null_checkpoint_id_r},
+        ).fetchone()
+
+    assert row is not None, (
+        f"Checkpoint {context.set_null_checkpoint_id_r!r} missing after resource deletion"
+    )
+    assert row[0] is None, (
+        f"Expected checkpoint resource_id to be NULL after parent deletion, "
+        f"got {row[0]!r}"
+    )

features/steps/db_schema_link_type_steps.py

@@ -0,0 +1,432 @@
+"""Step definitions for db_migration_lifecycle.feature — link type and migration.
+
+Link-type acceptance/rejection, migration idempotency (else-branch), orphan
+cleanup during migration, and downgrade verification for the m4_004 schema
+parity migration.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from alembic import command
+from behave import given, then, when
+from sqlalchemy import inspect as sa_inspect
+from sqlalchemy import text
+from sqlalchemy.exc import IntegrityError
+
+from cleveragents.infrastructure.database.migration_runner import MigrationRunner
+from features.steps.db_schema_parity_steps import _insert_test_resource
+
+# ---------------------------------------------------------------------------
+# Given — migration idempotency (link_type pre-exists)
+# ---------------------------------------------------------------------------
+
+
+@given("resource_links already has a link_type column without CHECK constraint")
+def step_add_link_type_without_check(context: Any) -> None:
+    """Add a bare link_type column so the migration else-branch is exercised."""
+    with context.engine.begin() as conn:
+        conn.execute(
+            text(
+                "ALTER TABLE resource_links "
+                "ADD COLUMN link_type TEXT DEFAULT 'contains'"
+            )
+        )
+
+
+@given("resource_links already has a link_type column with a non-contains default")
+def step_add_link_type_wrong_default(context: Any) -> None:
+    """Add a link_type column with wrong default so the default-fix path runs."""
+    with context.engine.begin() as conn:
+        conn.execute(
+            text("ALTER TABLE resource_links ADD COLUMN link_type TEXT DEFAULT 'other'")
+        )
+
+
+# ---------------------------------------------------------------------------
+# Given/When/Then — orphan cleanup during migration
+# ---------------------------------------------------------------------------
+
+
+@given('migrations applied up to "{revision}"')
+def step_migrations_up_to(context: Any, revision: str) -> None:
+    runner = MigrationRunner(context.db_url)
+    with context.engine.connect() as conn:
+        runner.alembic_cfg.attributes["connection"] = conn
+        try:
+            command.upgrade(runner.alembic_cfg, revision)
+            conn.commit()
+        finally:
+            runner.alembic_cfg.attributes.pop("connection", None)
+    context.runner = runner
+
+
+@given("checkpoint_metadata contains orphan decision and resource references")
+def step_insert_orphan_checkpoint_refs(context: Any) -> None:
+    action_name = "local/orphan-cleanup-action"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FD0"
+    checkpoint_id_1 = "01ARZ3NDEKTSV4RRFFQ69G5FD1"
+    checkpoint_id_2 = "01ARZ3NDEKTSV4RRFFQ69G5FD2"
+
+    with context.engine.begin() as conn:
+        conn.execute(
+            text(
+                """
+                INSERT INTO actions (
+                    namespaced_name, namespace, name, description,
+                    definition_of_done, strategy_actor, execution_actor,
+                    created_at, updated_at
+                ) VALUES (
+                    :namespaced_name, :namespace, :name, :description,
+                    :definition_of_done, :strategy_actor, :execution_actor,
+                    :created_at, :updated_at
+                )
+                """
+            ),
+            {
+                "namespaced_name": action_name,
+                "namespace": "local",
+                "name": "orphan-cleanup-action",
+                "description": "test action",
+                "definition_of_done": "test dod",
+                "strategy_actor": "local/strategy",
+                "execution_actor": "local/execution",
+                "created_at": "2026-01-01T00:00:00",
+                "updated_at": "2026-01-01T00:00:00",
+            },
+        )
+        # Detect whether root_plan_id column exists (added by
+        # m8_001_align_plans_schema which may not yet be applied when
+        # this step runs at the m4_003 migration state).
+        plan_columns = {
+            col["name"] for col in sa_inspect(context.engine).get_columns("v3_plans")
+        }
+        plan_cols = (
+            "plan_id, action_name, namespaced_name, namespace,"
+            " description, created_at, updated_at"
+        )
+        plan_vals = (
+            ":plan_id, :action_name, :namespaced_name, :namespace,"
+            " :description, :created_at, :updated_at"
+        )
+        plan_params: dict[str, str] = {
+            "plan_id": plan_id,
+            "action_name": action_name,
+            "namespaced_name": "local/orphan-plan",
+            "namespace": "local",
+            "description": "test plan for orphan cleanup",
+            "created_at": "2026-01-01T00:00:00",
+            "updated_at": "2026-01-01T00:00:00",
+        }
+        if "root_plan_id" in plan_columns:
+            plan_cols = (
+                "plan_id, root_plan_id, action_name,"
+                " namespaced_name, namespace,"
+                " description, created_at, updated_at"
+            )
+            plan_vals = (
+                ":plan_id, :root_plan_id, :action_name,"
+                " :namespaced_name, :namespace,"
+                " :description, :created_at, :updated_at"
+            )
+            plan_params["root_plan_id"] = plan_id
+        conn.execute(
+            text(f"INSERT INTO v3_plans ({plan_cols}) VALUES ({plan_vals})"),
+            plan_params,
+        )
+        # Checkpoint with orphan decision_id
+        conn.execute(
+            text(
+                """
+                INSERT INTO checkpoint_metadata (
+                    checkpoint_id, plan_id, decision_id,
+                    checkpoint_type, sandbox_ref,
+                    filesystem_path, created_at
+                ) VALUES (
+                    :checkpoint_id, :plan_id, :decision_id,
+                    :checkpoint_type, :sandbox_ref,
+                    :filesystem_path, :created_at
+                )
+                """
+            ),
+            {
+                "checkpoint_id": checkpoint_id_1,
+                "plan_id": plan_id,
+                "decision_id": "ORPHAN_DECISION_ID_NONEXIST",
+                "checkpoint_type": "manual",
+                "sandbox_ref": "test-ref",
+                "filesystem_path": "",
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+        # Checkpoint with orphan resource_id
+        conn.execute(
+            text(
+                """
+                INSERT INTO checkpoint_metadata (
+                    checkpoint_id, plan_id, resource_id,
+                    checkpoint_type, sandbox_ref,
+                    filesystem_path, created_at
+                ) VALUES (
+                    :checkpoint_id, :plan_id, :resource_id,
+                    :checkpoint_type, :sandbox_ref,
+                    :filesystem_path, :created_at
+                )
+                """
+            ),
+            {
+                "checkpoint_id": checkpoint_id_2,
+                "plan_id": plan_id,
+                "resource_id": "ORPHAN_RESOURCE_ID_NONEXIST",
+                "checkpoint_type": "manual",
+                "sandbox_ref": "test-ref",
+                "filesystem_path": "",
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+    context.orphan_plan_id = plan_id
+    context.orphan_checkpoint_id_1 = checkpoint_id_1
+    context.orphan_checkpoint_id_2 = checkpoint_id_2
+
+
+@when("I upgrade to the next migration revision")
+def step_upgrade_one_revision(context: Any) -> None:
+    runner = context.runner
+    with context.engine.connect() as conn:
+        runner.alembic_cfg.attributes["connection"] = conn
+        try:
+            # Target m4_004 explicitly because m4_003 has multiple
+            # child branches (m4_004, m5_001, m8_001_*) and "+1"
+            # would cause an "Ambiguous walk" error.
+            command.upgrade(
+                runner.alembic_cfg,
+                "m4_004_schema_parity_resource_decision_checkpoint",
+            )
+            conn.commit()
+        finally:
+            runner.alembic_cfg.attributes.pop("connection", None)
+
+
+@then("the orphan decision_id values should be NULL")
+def step_orphan_decision_id_null(context: Any) -> None:
+    with context.engine.connect() as conn:
+        row = conn.execute(
+            text(
+                "SELECT decision_id FROM checkpoint_metadata WHERE checkpoint_id = :cid"
+            ),
+            {"cid": context.orphan_checkpoint_id_1},
+        ).fetchone()
+    assert row is not None, (
+        f"Checkpoint {context.orphan_checkpoint_id_1!r} missing after migration"
+    )
+    assert row[0] is None, f"Expected orphan decision_id to be NULL, got {row[0]!r}"
+
+
+@then("the orphan resource_id values should be NULL")
+def step_orphan_resource_id_null(context: Any) -> None:
+    with context.engine.connect() as conn:
+        row = conn.execute(
+            text(
+                "SELECT resource_id FROM checkpoint_metadata WHERE checkpoint_id = :cid"
+            ),
+            {"cid": context.orphan_checkpoint_id_2},
+        ).fetchone()
+    assert row is not None, (
+        f"Checkpoint {context.orphan_checkpoint_id_2!r} missing after migration"
+    )
+    assert row[0] is None, f"Expected orphan resource_id to be NULL, got {row[0]!r}"
+
+
+# ---------------------------------------------------------------------------
+# Then — link_type acceptance/rejection
+# ---------------------------------------------------------------------------
+
+
+@then('resource_links should accept link_type "{link_type_value}"')
+def step_resource_links_accepts_link_type(context: Any, link_type_value: str) -> None:
+    with context.engine.begin() as conn:
+        parent_id = f"01LINKTYPE_{link_type_value.upper()[:6]}P"
+        child_id = f"01LINKTYPE_{link_type_value.upper()[:6]}C"
+
+        for rid in (parent_id, child_id):
+            _insert_test_resource(conn, rid)
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO resource_links (parent_id, child_id, link_type, created_at)
+                VALUES (:parent_id, :child_id, :link_type, :created_at)
+                """
+            ),
+            {
+                "parent_id": parent_id,
+                "child_id": child_id,
+                "link_type": link_type_value,
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+
+@then("resource_links should reject NULL link_type")
+def step_resource_links_rejects_null_link_type(context: Any) -> None:
+    with context.engine.begin() as conn:
+        parent_id = "01LINKNULLP_TESTPARENT"
+        child_id = "01LINKNULLC_TESTCHILD0"
+
+        for rid in (parent_id, child_id):
+            _insert_test_resource(conn, rid)
+
+        try:
+            conn.execute(
+                text(
+                    """
+                    INSERT INTO resource_links (parent_id, child_id, link_type, created_at)
+                    VALUES (:parent_id, :child_id, NULL, :created_at)
+                    """
+                ),
+                {
+                    "parent_id": parent_id,
+                    "child_id": child_id,
+                    "created_at": "2026-01-01T00:00:00",
+                },
+            )
+        except IntegrityError:
+            return
+
+    raise AssertionError("resource_links accepted NULL link_type")
+
+
+@then('resource_links should reject link_type "{link_type_value}"')
+def step_resource_links_rejects_link_type(context: Any, link_type_value: str) -> None:
+    with context.engine.begin() as conn:
+        parent_id = "01LINKTYPE_INVALIDP_TEST"
+        child_id = "01LINKTYPE_INVALIDC_TEST"
+
+        for rid in (parent_id, child_id):
+            _insert_test_resource(conn, rid)
+
+        try:
+            conn.execute(
+                text(
+                    """
+                    INSERT INTO resource_links (parent_id, child_id, link_type, created_at)
+                    VALUES (:parent_id, :child_id, :link_type, :created_at)
+                    """
+                ),
+                {
+                    "parent_id": parent_id,
+                    "child_id": child_id,
+                    "link_type": link_type_value,
+                    "created_at": "2026-01-01T00:00:00",
+                },
+            )
+        except IntegrityError:
+            return
+
+    raise AssertionError(
+        f"resource_links accepted invalid link_type {link_type_value!r}"
+    )
+
+
+@then("resource_links should reject empty string link_type")
+def step_resource_links_rejects_empty_link_type(context: Any) -> None:
+    with context.engine.begin() as conn:
+        parent_id = "01LINKTYPE_EMPTYP_TESTXX"
+        child_id = "01LINKTYPE_EMPTYC_TESTXX"
+
+        for rid in (parent_id, child_id):
+            _insert_test_resource(conn, rid)
+
+        try:
+            conn.execute(
+                text(
+                    """
+                    INSERT INTO resource_links (parent_id, child_id, link_type, created_at)
+                    VALUES (:parent_id, :child_id, :link_type, :created_at)
+                    """
+                ),
+                {
+                    "parent_id": parent_id,
+                    "child_id": child_id,
+                    "link_type": "",
+                    "created_at": "2026-01-01T00:00:00",
+                },
+            )
+        except IntegrityError:
+            return
+
+    raise AssertionError("resource_links accepted empty string link_type")
+
+
+# ---------------------------------------------------------------------------
+# When/Then — downgrade verification
+# ---------------------------------------------------------------------------
+
+
+@when('I downgrade to revision "{revision}"')
+def step_downgrade_to_specific_revision(context: Any, revision: str) -> None:
+    if not hasattr(context, "runner"):
+        runner = MigrationRunner(context.db_url)
+        runner.run_migrations(engine=context.engine)
+        context.runner = runner
+
+    runner = context.runner
+    with context.engine.connect() as conn:
+        runner.alembic_cfg.attributes["connection"] = conn
+        try:
+            command.downgrade(runner.alembic_cfg, revision)
+            conn.commit()
+        finally:
+            runner.alembic_cfg.attributes.pop("connection", None)
+
+
+@then('the "resource_links" table should not include "{column_name}"')
+def step_table_should_not_include_column(context: Any, column_name: str) -> None:
+    inspector = sa_inspect(context.engine)
+    columns = {col["name"] for col in inspector.get_columns("resource_links")}
+    assert column_name not in columns, (
+        f"Expected column {column_name!r} to be absent from resource_links, "
+        f"but found it in {columns!r}"
+    )
+
+
+@then('the "decisions" table should not have index "{index_name}"')
+def step_decisions_should_not_have_index(context: Any, index_name: str) -> None:
+    inspector = sa_inspect(context.engine)
+    indexes = {idx["name"] for idx in inspector.get_indexes("decisions")}
+    assert index_name not in indexes, (
+        f"Expected index {index_name!r} to be absent from decisions, "
+        f"but found it in {indexes!r}"
+    )
+
+
+@then('checkpoint_metadata should not have foreign key "{fk_name}"')
+def step_checkpoint_should_not_have_fk(context: Any, fk_name: str) -> None:
+    inspector = sa_inspect(context.engine)
+    fks = inspector.get_foreign_keys("checkpoint_metadata")
+    fk_names = {fk.get("name") for fk in fks}
+    assert fk_name not in fk_names, (
+        f"Expected FK {fk_name!r} to be absent from checkpoint_metadata, "
+        f"but found it in {fk_names!r}"
+    )
+
+
+@then("checkpoint_metadata should not have SQLite triggers for FK enforcement")
+def step_checkpoint_no_fk_triggers(context: Any) -> None:
+    if context.engine.dialect.name != "sqlite":
+        return  # Triggers are SQLite-specific; skip on other dialects.
+
+    with context.engine.connect() as conn:
+        rows = conn.execute(
+            text(
+                "SELECT name FROM sqlite_master "
+                "WHERE type = 'trigger' AND name LIKE 'trg_checkpoint_metadata_%'"
+            )
+        ).fetchall()
+
+    trigger_names = [row[0] for row in rows]
+    assert not trigger_names, (
+        f"Expected no trg_checkpoint_metadata_* triggers after downgrade, "
+        f"found: {trigger_names!r}"
+    )

features/steps/db_schema_parity_steps.py

@@ -0,0 +1,427 @@
+"""Step definitions for db_migration_lifecycle.feature — schema parity.
+
+FK structure verification, orphan rejection, valid references, and
+partial index checks for the ``resource_links``, ``checkpoint_metadata``,
+and ``decisions`` tables.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from behave import then
+from sqlalchemy import inspect as sa_inspect
+from sqlalchemy import text
+from sqlalchemy.exc import IntegrityError
+
+# ---------------------------------------------------------------------------
+# Shared test-data helpers
+# ---------------------------------------------------------------------------
+
+
+def _ensure_test_action_and_plan(conn: Any, action_name: str, plan_id: str) -> None:
+    """Insert prerequisite action and plan rows if absent."""
+    existing_action = conn.execute(
+        text("SELECT 1 FROM actions WHERE namespaced_name = :n"),
+        {"n": action_name},
+    ).fetchone()
+    existing_plan = conn.execute(
+        text("SELECT 1 FROM v3_plans WHERE plan_id = :pid"),
+        {"pid": plan_id},
+    ).fetchone()
+    if existing_action is not None and existing_plan is not None:
+        return
+
+    if existing_action is None:
+        conn.execute(
+            text(
+                """
+                INSERT INTO actions (
+                    namespaced_name, namespace, name, description,
+                    definition_of_done, strategy_actor, execution_actor,
+                    created_at, updated_at
+                ) VALUES (
+                    :namespaced_name, :namespace, :name, :description,
+                    :definition_of_done, :strategy_actor, :execution_actor,
+                    :created_at, :updated_at
+                )
+                """
+            ),
+            {
+                "namespaced_name": action_name,
+                "namespace": "local",
+                "name": action_name.split("/")[-1],
+                "description": "test action",
+                "definition_of_done": "test dod",
+                "strategy_actor": "local/strategy",
+                "execution_actor": "local/execution",
+                "created_at": "2026-01-01T00:00:00",
+                "updated_at": "2026-01-01T00:00:00",
+            },
+        )
+
+    if existing_plan is None:
+        conn.execute(
+            text(
+                """
+                INSERT INTO v3_plans (
+                    plan_id, root_plan_id, action_name,
+                    namespaced_name, namespace,
+                    description, created_at, updated_at
+                ) VALUES (
+                    :plan_id, :root_plan_id, :action_name,
+                    :namespaced_name, :namespace,
+                    :description, :created_at, :updated_at
+                )
+                """
+            ),
+            {
+                "plan_id": plan_id,
+                "root_plan_id": plan_id,
+                "action_name": action_name,
+                "namespaced_name": "local/test-plan-fk",
+                "namespace": "local",
+                "description": "test plan",
+                "created_at": "2026-01-01T00:00:00",
+                "updated_at": "2026-01-01T00:00:00",
+            },
+        )
+
+
+def _insert_test_resource(
+    conn: Any,
+    resource_id: str,
+    type_name: str = "git-checkout",
+    resource_kind: str = "physical",
+) -> None:
+    """Insert a test resource row, handling the optional namespaced_name column."""
+    resource_columns = {
+        col["name"] for col in sa_inspect(conn).get_columns("resources")
+    }
+    cols = "resource_id, type_name, resource_kind, created_at, updated_at"
+    vals = ":resource_id, :type_name, :resource_kind, :created_at, :updated_at"
+    params: dict[str, str] = {
+        "resource_id": resource_id,
+        "type_name": type_name,
+        "resource_kind": resource_kind,
+        "created_at": "2026-01-01T00:00:00",
+        "updated_at": "2026-01-01T00:00:00",
+    }
+    if "namespaced_name" in resource_columns:
+        cols = (
+            "resource_id, namespaced_name, type_name,"
+            " resource_kind, created_at, updated_at"
+        )
+        vals = (
+            ":resource_id, :namespaced_name, :type_name,"
+            " :resource_kind, :created_at, :updated_at"
+        )
+        params["namespaced_name"] = f"local/{resource_id}"
+    conn.execute(text(f"INSERT INTO resources ({cols}) VALUES ({vals})"), params)
+
+
+# ---------------------------------------------------------------------------
+# Then — link_type default verification
+# ---------------------------------------------------------------------------
+
+
+@then('the "resource_links" table should include "link_type" with default "contains"')
+def step_resource_links_has_link_type_default(context: Any) -> None:
+    inspector = sa_inspect(context.engine)
+    columns = inspector.get_columns("resource_links")
+    link_type = next(
+        (column for column in columns if column["name"] == "link_type"), None
+    )
+    assert link_type is not None, "resource_links.link_type column is missing"
+
+    default = str(link_type.get("default") or "").lower()
+    assert "contains" in default, (
+        "Expected resource_links.link_type default to include 'contains', "
+        f"got: {link_type.get('default')!r}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Then — checkpoint FK structure verification
+# ---------------------------------------------------------------------------
+
+
+@then("checkpoint_metadata should enforce decision and resource foreign keys")
+def step_checkpoint_metadata_foreign_keys(context: Any) -> None:
+    inspector = sa_inspect(context.engine)
+    foreign_keys = inspector.get_foreign_keys("checkpoint_metadata")
+    signatures = {
+        (
+            tuple(fk.get("constrained_columns") or []),
+            fk.get("referred_table"),
+            tuple(fk.get("referred_columns") or []),
+        )
+        for fk in foreign_keys
+    }
+
+    decision_fk = (("decision_id",), "decisions", ("decision_id",))
+    resource_fk = (("resource_id",), "resources", ("resource_id",))
+
+    assert decision_fk in signatures, (
+        "Missing checkpoint_metadata foreign key for decision_id -> "
+        "decisions.decision_id"
+    )
+    assert resource_fk in signatures, (
+        "Missing checkpoint_metadata foreign key for resource_id -> "
+        "resources.resource_id"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Then — orphan FK rejection (combined)
+# ---------------------------------------------------------------------------
+
+
+@then("checkpoint_metadata foreign keys should reject orphan references")
+def step_checkpoint_metadata_foreign_keys_reject_orphans(context: Any) -> None:
+    action_name = "local/test-action"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FAV"
+
+    with context.engine.begin() as conn:
+        _ensure_test_action_and_plan(conn, action_name, plan_id)
+
+        try:
+            conn.execute(
+                text(
+                    """
+                    INSERT INTO checkpoint_metadata (
+                        checkpoint_id,
+                        plan_id,
+                        decision_id,
+                        checkpoint_type,
+                        resource_id,
+                        sandbox_ref,
+                        filesystem_path,
+                        created_at
+                    ) VALUES (
+                        :checkpoint_id,
+                        :plan_id,
+                        :decision_id,
+                        :checkpoint_type,
+                        :resource_id,
+                        :sandbox_ref,
+                        :filesystem_path,
+                        :created_at
+                    )
+                    """
+                ),
+                {
+                    "checkpoint_id": "01ARZ3NDEKTSV4RRFFQ69G5FAW",
+                    "plan_id": plan_id,
+                    "decision_id": "01ARZ3NDEKTSV4RRFFQ69G5FAX",
+                    "checkpoint_type": "manual",
+                    "resource_id": "01ARZ3NDEKTSV4RRFFQ69G5FAY",
+                    "sandbox_ref": "test-ref",
+                    "filesystem_path": "",
+                    "created_at": "2026-01-01T00:00:00",
+                },
+            )
+        except IntegrityError:
+            return
+
+    raise AssertionError(
+        "checkpoint_metadata accepted orphan decision/resource references"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Then — independent orphan FK rejection
+# ---------------------------------------------------------------------------
+
+
+@then("checkpoint_metadata should reject orphan decision_id independently")
+def step_checkpoint_reject_orphan_decision_only(context: Any) -> None:
+    action_name = "local/test-action-fk-d"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FB0"
+
+    with context.engine.begin() as conn:
+        _ensure_test_action_and_plan(conn, action_name, plan_id)
+
+        try:
+            conn.execute(
+                text(
+                    """
+                    INSERT INTO checkpoint_metadata (
+                        checkpoint_id, plan_id, decision_id,
+                        checkpoint_type, sandbox_ref,
+                        filesystem_path, created_at
+                    ) VALUES (
+                        :checkpoint_id, :plan_id, :decision_id,
+                        :checkpoint_type, :sandbox_ref,
+                        :filesystem_path, :created_at
+                    )
+                    """
+                ),
+                {
+                    "checkpoint_id": "01ARZ3NDEKTSV4RRFFQ69G5FB1",
+                    "plan_id": plan_id,
+                    "decision_id": "01ARZ3NDEKTSV4RRFFQ69G5FB2",
+                    "checkpoint_type": "manual",
+                    "sandbox_ref": "test-ref",
+                    "filesystem_path": "",
+                    "created_at": "2026-01-01T00:00:00",
+                },
+            )
+        except IntegrityError:
+            return
+
+    raise AssertionError(
+        "checkpoint_metadata accepted orphan decision_id with NULL resource_id"
+    )
+
+
+@then("checkpoint_metadata should reject orphan resource_id independently")
+def step_checkpoint_reject_orphan_resource_only(context: Any) -> None:
+    action_name = "local/test-action-fk-r"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FB3"
+
+    with context.engine.begin() as conn:
+        _ensure_test_action_and_plan(conn, action_name, plan_id)
+
+        try:
+            conn.execute(
+                text(
+                    """
+                    INSERT INTO checkpoint_metadata (
+                        checkpoint_id, plan_id, resource_id,
+                        checkpoint_type, sandbox_ref,
+                        filesystem_path, created_at
+                    ) VALUES (
+                        :checkpoint_id, :plan_id, :resource_id,
+                        :checkpoint_type, :sandbox_ref,
+                        :filesystem_path, :created_at
+                    )
+                    """
+                ),
+                {
+                    "checkpoint_id": "01ARZ3NDEKTSV4RRFFQ69G5FB4",
+                    "plan_id": plan_id,
+                    "resource_id": "01ARZ3NDEKTSV4RRFFQ69G5FB5",
+                    "checkpoint_type": "manual",
+                    "sandbox_ref": "test-ref",
+                    "filesystem_path": "",
+                    "created_at": "2026-01-01T00:00:00",
+                },
+            )
+        except IntegrityError:
+            return
+
+    raise AssertionError(
+        "checkpoint_metadata accepted orphan resource_id with NULL decision_id"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Then — valid FK acceptance
+# ---------------------------------------------------------------------------
+
+
+@then("checkpoint_metadata should accept valid decision and resource references")
+def step_checkpoint_accept_valid_references(context: Any) -> None:
+    action_name = "local/test-action-fk-v"
+    plan_id = "01ARZ3NDEKTSV4RRFFQ69G5FB6"
+    decision_id = "01ARZ3NDEKTSV4RRFFQ69G5FB7"
+    resource_id = "01ARZ3NDEKTSV4RRFFQ69G5FB8"
+
+    with context.engine.begin() as conn:
+        _ensure_test_action_and_plan(conn, action_name, plan_id)
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO decisions (
+                    decision_id, plan_id, decision_type, question,
+                    chosen_option, context_snapshot_json, sequence_number,
+                    created_at
+                ) VALUES (
+                    :decision_id, :plan_id, :decision_type, :question,
+                    :chosen_option, :context_snapshot_json, :sequence_number,
+                    :created_at
+                )
+                """
+            ),
+            {
+                "decision_id": decision_id,
+                "plan_id": plan_id,
+                "decision_type": "strategy_choice",
+                "question": "test question",
+                "chosen_option": "test option",
+                "context_snapshot_json": "{}",
+                "sequence_number": 1,
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+        _insert_test_resource(conn, resource_id)
+
+        conn.execute(
+            text(
+                """
+                INSERT INTO checkpoint_metadata (
+                    checkpoint_id, plan_id, decision_id,
+                    checkpoint_type, resource_id, sandbox_ref,
+                    filesystem_path, created_at
+                ) VALUES (
+                    :checkpoint_id, :plan_id, :decision_id,
+                    :checkpoint_type, :resource_id, :sandbox_ref,
+                    :filesystem_path, :created_at
+                )
+                """
+            ),
+            {
+                "checkpoint_id": "01ARZ3NDEKTSV4RRFFQ69G5FB9",
+                "plan_id": plan_id,
+                "decision_id": decision_id,
+                "checkpoint_type": "manual",
+                "resource_id": resource_id,
+                "sandbox_ref": "test-ref",
+                "filesystem_path": "",
+                "created_at": "2026-01-01T00:00:00",
+            },
+        )
+
+
+# ---------------------------------------------------------------------------
+# Then — partial index verification
+# ---------------------------------------------------------------------------
+
+
+@then(
+    'the "decisions" table should have partial index "{index_name}" on "{column_name}"'
+)
+def step_decisions_has_partial_index(
+    context: Any,
+    index_name: str,
+    column_name: str,
+) -> None:
+    inspector = sa_inspect(context.engine)
+    indexes = inspector.get_indexes("decisions")
+    index = next((idx for idx in indexes if idx.get("name") == index_name), None)
+    assert index is not None, f"Index {index_name} not found on decisions"
+
+    columns = list(index.get("column_names") or [])
+    assert columns == [column_name], (
+        f"Index {index_name} expected on [{column_name!r}], got {columns!r}"
+    )
+
+    # Verify partial WHERE clause via sqlite_master (SQLite-only).
+    if context.engine.dialect.name == "sqlite":
+        with context.engine.connect() as conn:
+            row = conn.execute(
+                text(
+                    "SELECT sql FROM sqlite_master "
+                    "WHERE type = 'index' AND name = :index_name"
+                ),
+                {"index_name": index_name},
+            ).fetchone()
+
+        assert row is not None, f"sqlite_master entry missing for index {index_name}"
+        sql = str(row[0] or "").lower()
+        assert "where superseded_by is not null" in sql, (
+            f"Expected partial WHERE clause on {index_name}, got SQL: {row[0]!r}"
+        )

robot/helper_schema_parity_migration.py

@@ -0,0 +1,443 @@
+"""Robot helper for schema-parity migration verification.
+
+Checks the migration-produced SQLite schema for:
+
+1. ``resource_links.link_type`` defaulting to ``contains``.
+2. ``checkpoint_metadata`` foreign keys to ``decisions`` and ``resources``.
+3. ``idx_decisions_superseded`` partial index with
+   ``WHERE superseded_by IS NOT NULL``.
+4. Runtime SQLite foreign key enforcement for ``checkpoint_metadata``.
+
+Subcommands (each independent for isolated failure reporting):
+
+- ``schema-parity-link-type``: Verifies resource_links.link_type column
+  and default.
+- ``schema-parity-fks``: Verifies checkpoint_metadata FK constraints and
+  runtime enforcement (orphan rejection + positive path).
+- ``schema-parity-index``: Verifies idx_decisions_superseded partial index.
+"""
+
+from __future__ import annotations
+
+import sys
+import tempfile
+from collections.abc import Callable
+from pathlib import Path
+from typing import Any
+
+from sqlalchemy import create_engine, inspect, text
+from sqlalchemy.exc import IntegrityError
+
+from cleveragents.infrastructure.database.migration_runner import MigrationRunner
+
+
+def _cleanup_db_files(path: str) -> None:
+    Path(path).unlink(missing_ok=True)
+    Path(f"{path}-wal").unlink(missing_ok=True)
+    Path(f"{path}-shm").unlink(missing_ok=True)
+
+
+def _setup_migrated_db() -> tuple[Any, Any, str]:
+    """Create a temp DB, run migrations, return (engine, inspector, db_path)."""
+    with tempfile.NamedTemporaryFile(suffix=".db", delete=False) as tmp_file:
+        db_path = tmp_file.name
+    db_url = f"sqlite:///{db_path}"
+
+    MigrationRunner(db_url).init_or_upgrade()
+    engine = create_engine(db_url, connect_args={"check_same_thread": False})
+    inspector = inspect(engine)
+    return engine, inspector, db_path
+
+
+def _teardown_db(engine: Any, db_path: str) -> None:
+    if engine is not None:
+        engine.dispose()
+    _cleanup_db_files(db_path)
+
+
+def _ensure_test_action_and_plan(conn: Any) -> None:
+    """Insert prerequisite action and plan rows for FK tests.
+
+    .. note::
+        This function is **not idempotent** — it performs blind INSERTs
+        without existence checks.  It is safe only when called against a
+        freshly created database (as ``_setup_migrated_db`` provides).
+        If idempotent behaviour is needed, see the Behave counterpart in
+        ``features/steps/db_schema_parity_steps.py``.
+    """
+    conn.execute(
+        text(
+            """
+            INSERT INTO actions (
+                namespaced_name, namespace, name, description,
+                definition_of_done, strategy_actor, execution_actor,
+                created_at, updated_at
+            ) VALUES (
+                :namespaced_name, :namespace, :name, :description,
+                :definition_of_done, :strategy_actor, :execution_actor,
+                :created_at, :updated_at
+            )
+            """
+        ),
+        {
+            "namespaced_name": "local/test-action",
+            "namespace": "local",
+            "name": "test-action",
+            "description": "test action",
+            "definition_of_done": "test dod",
+            "strategy_actor": "local/strategy",
+            "execution_actor": "local/execution",
+            "created_at": "2026-01-01T00:00:00",
+            "updated_at": "2026-01-01T00:00:00",
+        },
+    )
+    conn.execute(
+        text(
+            """
+            INSERT INTO v3_plans (
+                plan_id, root_plan_id, action_name,
+                namespaced_name, namespace,
+                description, created_at, updated_at
+            ) VALUES (
+                :plan_id, :root_plan_id, :action_name,
+                :namespaced_name, :namespace,
+                :description, :created_at, :updated_at
+            )
+            """
+        ),
+        {
+            "plan_id": "01ARZ3NDEKTSV4RRFFQ69G5FAV",
+            "root_plan_id": "01ARZ3NDEKTSV4RRFFQ69G5FAV",
+            "action_name": "local/test-action",
+            "namespaced_name": "local/test-plan",
+            "namespace": "local",
+            "description": "test plan",
+            "created_at": "2026-01-01T00:00:00",
+            "updated_at": "2026-01-01T00:00:00",
+        },
+    )
+
+
+# ---------------------------------------------------------------------------
+# Subcommand: schema-parity-link-type
+# ---------------------------------------------------------------------------
+
+
+def _schema_parity_link_type() -> None:
+    engine, inspector, db_path = _setup_migrated_db()
+    try:
+        resource_link_columns = inspector.get_columns("resource_links")
+        link_type = next(
+            (
+                column
+                for column in resource_link_columns
+                if column["name"] == "link_type"
+            ),
+            None,
+        )
+        assert link_type is not None, "resource_links.link_type column is missing"
+        default = str(link_type.get("default") or "").lower()
+        assert "contains" in default, (
+            "resource_links.link_type default must include 'contains', "
+            f"got {link_type.get('default')!r}"
+        )
+
+        print("schema-parity-link-type-ok")
+    finally:
+        _teardown_db(engine, db_path)
+
+
+# ---------------------------------------------------------------------------
+# Subcommand: schema-parity-fks
+# ---------------------------------------------------------------------------
+
+
+def _schema_parity_fks() -> None:
+    engine, inspector, db_path = _setup_migrated_db()
+    try:
+        checkpoint_fks = inspector.get_foreign_keys("checkpoint_metadata")
+        signatures = {
+            (
+                tuple(fk.get("constrained_columns") or []),
+                fk.get("referred_table"),
+                tuple(fk.get("referred_columns") or []),
+            )
+            for fk in checkpoint_fks
+        }
+
+        assert (("decision_id",), "decisions", ("decision_id",)) in signatures, (
+            "Missing checkpoint_metadata FK decision_id -> decisions.decision_id"
+        )
+        assert (("resource_id",), "resources", ("resource_id",)) in signatures, (
+            "Missing checkpoint_metadata FK resource_id -> resources.resource_id"
+        )
+
+        with engine.begin() as conn:
+            _ensure_test_action_and_plan(conn)
+
+            try:
+                conn.execute(
+                    text(
+                        """
+                        INSERT INTO checkpoint_metadata (
+                            checkpoint_id, plan_id, decision_id,
+                            checkpoint_type, resource_id, sandbox_ref,
+                            filesystem_path, created_at
+                        ) VALUES (
+                            :checkpoint_id, :plan_id, :decision_id,
+                            :checkpoint_type, :resource_id, :sandbox_ref,
+                            :filesystem_path, :created_at
+                        )
+                        """
+                    ),
+                    {
+                        "checkpoint_id": "01ARZ3NDEKTSV4RRFFQ69G5FAW",
+                        "plan_id": "01ARZ3NDEKTSV4RRFFQ69G5FAV",
+                        "decision_id": "01ARZ3NDEKTSV4RRFFQ69G5FAX",
+                        "checkpoint_type": "manual",
+                        "resource_id": "01ARZ3NDEKTSV4RRFFQ69G5FAY",
+                        "sandbox_ref": "test-ref",
+                        "filesystem_path": "",
+                        "created_at": "2026-01-01T00:00:00",
+                    },
+                )
+            except IntegrityError:
+                pass
+            else:
+                raise AssertionError(
+                    "checkpoint_metadata accepted orphan decision/resource references"
+                )
+
+        # Verify each FK independently: orphan decision_id only
+        with engine.begin() as conn:
+            try:
+                conn.execute(
+                    text(
+                        """
+                        INSERT INTO checkpoint_metadata (
+                            checkpoint_id, plan_id, decision_id,
+                            checkpoint_type, sandbox_ref,
+                            filesystem_path, created_at
+                        ) VALUES (
+                            :checkpoint_id, :plan_id, :decision_id,
+                            :checkpoint_type, :sandbox_ref,
+                            :filesystem_path, :created_at
+                        )
+                        """
+                    ),
+                    {
+                        "checkpoint_id": "01ARZ3NDEKTSV4RRFFQ69G5FC0",
+                        "plan_id": "01ARZ3NDEKTSV4RRFFQ69G5FAV",
+                        "decision_id": "01ARZ3NDEKTSV4RRFFQ69G5FC1",
+                        "checkpoint_type": "manual",
+                        "sandbox_ref": "test-ref",
+                        "filesystem_path": "",
+                        "created_at": "2026-01-01T00:00:00",
+                    },
+                )
+            except IntegrityError:
+                pass
+            else:
+                raise AssertionError(
+                    "checkpoint_metadata accepted orphan decision_id independently"
+                )
+
+        # Verify each FK independently: orphan resource_id only
+        with engine.begin() as conn:
+            try:
+                conn.execute(
+                    text(
+                        """
+                        INSERT INTO checkpoint_metadata (
+                            checkpoint_id, plan_id, resource_id,
+                            checkpoint_type, sandbox_ref,
+                            filesystem_path, created_at
+                        ) VALUES (
+                            :checkpoint_id, :plan_id, :resource_id,
+                            :checkpoint_type, :sandbox_ref,
+                            :filesystem_path, :created_at
+                        )
+                        """
+                    ),
+                    {
+                        "checkpoint_id": "01ARZ3NDEKTSV4RRFFQ69G5FC2",
+                        "plan_id": "01ARZ3NDEKTSV4RRFFQ69G5FAV",
+                        "resource_id": "01ARZ3NDEKTSV4RRFFQ69G5FC3",
+                        "checkpoint_type": "manual",
+                        "sandbox_ref": "test-ref",
+                        "filesystem_path": "",
+                        "created_at": "2026-01-01T00:00:00",
+                    },
+                )
+            except IntegrityError:
+                pass
+            else:
+                raise AssertionError(
+                    "checkpoint_metadata accepted orphan resource_id independently"
+                )
+
+        # Positive test: valid FK references should be accepted
+        with engine.begin() as conn:
+            decision_id = "01ARZ3NDEKTSV4RRFFQ69G5FC4"
+            resource_id = "01ARZ3NDEKTSV4RRFFQ69G5FC5"
+
+            conn.execute(
+                text(
+                    """
+                    INSERT INTO decisions (
+                        decision_id, plan_id, decision_type, question,
+                        chosen_option, context_snapshot_json, sequence_number,
+                        created_at
+                    ) VALUES (
+                        :decision_id, :plan_id, :decision_type, :question,
+                        :chosen_option, :context_snapshot_json, :sequence_number,
+                        :created_at
+                    )
+                    """
+                ),
+                {
+                    "decision_id": decision_id,
+                    "plan_id": "01ARZ3NDEKTSV4RRFFQ69G5FAV",
+                    "decision_type": "strategy_choice",
+                    "question": "test question",
+                    "chosen_option": "test option",
+                    "context_snapshot_json": "{}",
+                    "sequence_number": 1,
+                    "created_at": "2026-01-01T00:00:00",
+                },
+            )
+
+            resource_columns = {
+                col["name"] for col in inspect(engine).get_columns("resources")
+            }
+            cols = "resource_id, type_name, resource_kind, created_at, updated_at"
+            vals = ":resource_id, :type_name, :resource_kind, :created_at, :updated_at"
+            res_params: dict[str, str] = {
+                "resource_id": resource_id,
+                "type_name": "git-checkout",
+                "resource_kind": "physical",
+                "created_at": "2026-01-01T00:00:00",
+                "updated_at": "2026-01-01T00:00:00",
+            }
+            if "namespaced_name" in resource_columns:
+                cols = (
+                    "resource_id, namespaced_name, type_name,"
+                    " resource_kind, created_at, updated_at"
+                )
+                vals = (
+                    ":resource_id, :namespaced_name, :type_name,"
+                    " :resource_kind, :created_at, :updated_at"
+                )
+                res_params["namespaced_name"] = f"local/{resource_id}"
+            conn.execute(
+                text(f"INSERT INTO resources ({cols}) VALUES ({vals})"),
+                res_params,
+            )
+
+            conn.execute(
+                text(
+                    """
+                    INSERT INTO checkpoint_metadata (
+                        checkpoint_id, plan_id, decision_id,
+                        checkpoint_type, resource_id, sandbox_ref,
+                        filesystem_path, created_at
+                    ) VALUES (
+                        :checkpoint_id, :plan_id, :decision_id,
+                        :checkpoint_type, :resource_id, :sandbox_ref,
+                        :filesystem_path, :created_at
+                    )
+                    """
+                ),
+                {
+                    "checkpoint_id": "01ARZ3NDEKTSV4RRFFQ69G5FC6",
+                    "plan_id": "01ARZ3NDEKTSV4RRFFQ69G5FAV",
+                    "decision_id": decision_id,
+                    "checkpoint_type": "manual",
+                    "resource_id": resource_id,
+                    "sandbox_ref": "test-ref",
+                    "filesystem_path": "",
+                    "created_at": "2026-01-01T00:00:00",
+                },
+            )
+
+        print("schema-parity-fks-ok")
+    finally:
+        _teardown_db(engine, db_path)
+
+
+# ---------------------------------------------------------------------------
+# Subcommand: schema-parity-index
+# ---------------------------------------------------------------------------
+
+
+def _schema_parity_index() -> None:
+    engine, inspector, db_path = _setup_migrated_db()
+    try:
+        decision_indexes = inspector.get_indexes("decisions")
+        partial_index = next(
+            (
+                index
+                for index in decision_indexes
+                if index.get("name") == "idx_decisions_superseded"
+            ),
+            None,
+        )
+        assert partial_index is not None, "idx_decisions_superseded is missing"
+        assert list(partial_index.get("column_names") or []) == ["superseded_by"], (
+            "idx_decisions_superseded must index decisions.superseded_by"
+        )
+
+        with engine.connect() as conn:
+            row = conn.execute(
+                text(
+                    "SELECT sql FROM sqlite_master "
+                    "WHERE type = 'index' AND name = :index_name"
+                ),
+                {"index_name": "idx_decisions_superseded"},
+            ).fetchone()
+
+        assert row is not None, "sqlite_master is missing idx_decisions_superseded"
+        sql = str(row[0] or "").lower()
+        assert "where superseded_by is not null" in sql, (
+            "idx_decisions_superseded is not a partial index"
+        )
+
+        print("schema-parity-index-ok")
+    finally:
+        _teardown_db(engine, db_path)
+
+
+# ---------------------------------------------------------------------------
+# Legacy combined subcommand (kept for backward compatibility)
+# ---------------------------------------------------------------------------
+
+
+def _schema_parity() -> None:
+    _schema_parity_link_type()
+    _schema_parity_fks()
+    _schema_parity_index()
+    print("schema-parity-ok")
+
+
+_COMMANDS: dict[str, Callable[[], None]] = {
+    "schema-parity": _schema_parity,
+    "schema-parity-link-type": _schema_parity_link_type,
+    "schema-parity-fks": _schema_parity_fks,
+    "schema-parity-index": _schema_parity_index,
+}
+
+
+def main() -> None:
+    if len(sys.argv) < 2:
+        raise SystemExit(f"Expected command argument. Valid: {', '.join(_COMMANDS)}")
+
+    command = sys.argv[1]
+    handler = _COMMANDS.get(command)
+    if handler is None:
+        raise SystemExit(f"Unknown command: {command}. Valid: {', '.join(_COMMANDS)}")
+
+    handler()
+
+
+if __name__ == "__main__":
+    main()

robot/schema_parity_migration.robot

@@ -0,0 +1,36 @@
+*** Settings ***
+Documentation     Integration checks for spec-parity schema constraints and indexes.
+...               Verifies migration output includes resource_links.link_type default,
+...               checkpoint_metadata FK enforcement, and decisions partial superseded index.
+...               Split into independent test cases for isolated failure reporting.
+Resource          ${CURDIR}/common.resource
+Suite Setup       Setup Test Environment
+Suite Teardown    Cleanup Test Environment
+
+*** Variables ***
+${HELPER_SCRIPT}    robot/helper_schema_parity_migration.py
+
+*** Test Cases ***
+Resource Links Link Type Default After Migration
+    [Documentation]    Validate resource_links.link_type column exists with default 'contains'.
+    [Tags]    database    migration    integration    link-type
+    ${result}=    Run Process    ${PYTHON}    ${HELPER_SCRIPT}    schema-parity-link-type    cwd=${WORKSPACE}
+    Log    ${result.stderr}
+    Should Be Equal As Integers    ${result.rc}    0
+    Should Contain    ${result.stdout}    schema-parity-link-type-ok
+
+Checkpoint Metadata FK Enforcement After Migration
+    [Documentation]    Validate checkpoint_metadata FK constraints and orphan rejection.
+    [Tags]    database    migration    integration    fks
+    ${result}=    Run Process    ${PYTHON}    ${HELPER_SCRIPT}    schema-parity-fks    cwd=${WORKSPACE}
+    Log    ${result.stderr}
+    Should Be Equal As Integers    ${result.rc}    0
+    Should Contain    ${result.stdout}    schema-parity-fks-ok
+
+Decisions Superseded Partial Index After Migration
+    [Documentation]    Validate idx_decisions_superseded partial index on decisions.superseded_by.
+    [Tags]    database    migration    integration    index
+    ${result}=    Run Process    ${PYTHON}    ${HELPER_SCRIPT}    schema-parity-index    cwd=${WORKSPACE}
+    Log    ${result.stderr}
+    Should Be Equal As Integers    ${result.rc}    0
+    Should Contain    ${result.stdout}    schema-parity-index-ok

src/cleveragents/infrastructure/database/migrations/versions/m4_004_schema_parity_resource_decision_checkpoint.py

@@ -0,0 +1,318 @@
+"""Align resource/decision/checkpoint schema gaps with specification DDL.
+
+Adds four spec-parity changes:
+
+1. ``resource_links.link_type`` with default ``'contains'``.
+2. Foreign keys on ``checkpoint_metadata.decision_id`` and
+   ``checkpoint_metadata.resource_id``.
+3. Partial index ``idx_decisions_superseded`` on
+   ``decisions.superseded_by`` where non-null.
+4. SQLite trigger guards to enforce checkpoint FK semantics at runtime.
+
+Revision ID: m4_004_schema_parity_resource_decision_checkpoint
+Revises: m4_003_plan_env_columns
+Create Date: 2026-03-26 00:00:00
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "m4_004_schema_parity_resource_decision_checkpoint"
+down_revision: str | Sequence[str] | None = "m4_003_plan_env_columns"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _inspector() -> sa.Inspector:
+    return sa.inspect(op.get_bind())
+
+
+def _create_sqlite_checkpoint_fk_triggers() -> None:
+    # NOTE: These triggers guard INSERT and UPDATE on checkpoint_metadata only.
+    # DELETE-direction enforcement (preventing deletion of a referenced decision
+    # or resource) relies on PRAGMA foreign_keys=ON at the connection level.
+    # This is a known limitation of trigger-based FK emulation on SQLite.
+    op.execute(
+        sa.text(
+            """
+            CREATE TRIGGER IF NOT EXISTS trg_checkpoint_metadata_decision_fk_insert
+            BEFORE INSERT ON checkpoint_metadata
+            FOR EACH ROW
+            WHEN NEW.decision_id IS NOT NULL
+              AND NOT EXISTS (
+                  SELECT 1 FROM decisions WHERE decision_id = NEW.decision_id
+              )
+            BEGIN
+                SELECT RAISE(ABORT, 'FOREIGN KEY constraint failed');
+            END;
+            """
+        )
+    )
+    op.execute(
+        sa.text(
+            """
+            CREATE TRIGGER IF NOT EXISTS trg_checkpoint_metadata_decision_fk_update
+            BEFORE UPDATE OF decision_id ON checkpoint_metadata
+            FOR EACH ROW
+            WHEN NEW.decision_id IS NOT NULL
+              AND NOT EXISTS (
+                  SELECT 1 FROM decisions WHERE decision_id = NEW.decision_id
+              )
+            BEGIN
+                SELECT RAISE(ABORT, 'FOREIGN KEY constraint failed');
+            END;
+            """
+        )
+    )
+    op.execute(
+        sa.text(
+            """
+            CREATE TRIGGER IF NOT EXISTS trg_checkpoint_metadata_resource_fk_insert
+            BEFORE INSERT ON checkpoint_metadata
+            FOR EACH ROW
+            WHEN NEW.resource_id IS NOT NULL
+              AND NOT EXISTS (
+                  SELECT 1 FROM resources WHERE resource_id = NEW.resource_id
+              )
+            BEGIN
+                SELECT RAISE(ABORT, 'FOREIGN KEY constraint failed');
+            END;
+            """
+        )
+    )
+    op.execute(
+        sa.text(
+            """
+            CREATE TRIGGER IF NOT EXISTS trg_checkpoint_metadata_resource_fk_update
+            BEFORE UPDATE OF resource_id ON checkpoint_metadata
+            FOR EACH ROW
+            WHEN NEW.resource_id IS NOT NULL
+              AND NOT EXISTS (
+                  SELECT 1 FROM resources WHERE resource_id = NEW.resource_id
+              )
+            BEGIN
+                SELECT RAISE(ABORT, 'FOREIGN KEY constraint failed');
+            END;
+            """
+        )
+    )
+
+
+def _drop_sqlite_checkpoint_fk_triggers() -> None:
+    op.execute(
+        sa.text("DROP TRIGGER IF EXISTS trg_checkpoint_metadata_decision_fk_insert")
+    )
+    op.execute(
+        sa.text("DROP TRIGGER IF EXISTS trg_checkpoint_metadata_decision_fk_update")
+    )
+    op.execute(
+        sa.text("DROP TRIGGER IF EXISTS trg_checkpoint_metadata_resource_fk_insert")
+    )
+    op.execute(
+        sa.text("DROP TRIGGER IF EXISTS trg_checkpoint_metadata_resource_fk_update")
+    )
+
+
+def upgrade() -> None:
+    """Apply spec-parity schema updates."""
+    inspector = _inspector()
+
+    resource_link_columns = {
+        column["name"] for column in inspector.get_columns("resource_links")
+    }
+    if "link_type" not in resource_link_columns:
+        with op.batch_alter_table("resource_links") as batch_op:
+            batch_op.add_column(
+                sa.Column(
+                    "link_type",
+                    sa.Text(),
+                    nullable=False,
+                    server_default=sa.text("'contains'"),
+                )
+            )
+            # NOTE: Spec DDL (line 45549) does not declare a CHECK on
+            # link_type; we add it to match the sibling resource_edges
+            # constraint and prevent invalid values at the DB level.
+            batch_op.create_check_constraint(
+                "ck_resource_links_link_type",
+                "link_type IN ('contains', 'references', 'derived_from')",
+            )
+    else:
+        # link_type column already exists — fix default and ensure CHECK
+        # constraint is present (guards against partial prior migration).
+        # NOTE: Spec DDL (line 45549) defines link_type as nullable
+        # (TEXT DEFAULT 'contains' without NOT NULL).  We deliberately
+        # enforce NOT NULL because a NULL link type is semantically
+        # meaningless and the sibling resource_edges table also requires
+        # a non-NULL link_type.
+        # NOTE: Column TYPE is not verified here.  On SQLite all text
+        # types are equivalent, so a prior VARCHAR(30) column works
+        # identically to TEXT.  On PostgreSQL, VARCHAR(30) != TEXT; if
+        # a future migration targets PostgreSQL this path should also
+        # include an ALTER COLUMN TYPE to sa.Text().
+        link_type_column = next(
+            column
+            for column in inspector.get_columns("resource_links")
+            if column["name"] == "link_type"
+        )
+        current_default = str(link_type_column.get("default") or "").lower()
+        needs_default_fix = "contains" not in current_default
+
+        check_constraints = inspector.get_check_constraints("resource_links")
+        has_link_type_ck = any(
+            ck.get("name") == "ck_resource_links_link_type" for ck in check_constraints
+        )
+
+        if needs_default_fix or not has_link_type_ck:
+            with op.batch_alter_table("resource_links") as batch_op:
+                if needs_default_fix:
+                    batch_op.alter_column(
+                        "link_type",
+                        server_default=sa.text("'contains'"),
+                    )
+                if not has_link_type_ck:
+                    batch_op.create_check_constraint(
+                        "ck_resource_links_link_type",
+                        "link_type IN ('contains', 'references', 'derived_from')",
+                    )
+
+    decision_indexes = {index["name"] for index in inspector.get_indexes("decisions")}
+    if "idx_decisions_superseded" not in decision_indexes:
+        op.create_index(
+            "idx_decisions_superseded",
+            "decisions",
+            ["superseded_by"],
+            unique=False,
+            postgresql_where=sa.text("superseded_by IS NOT NULL"),
+            sqlite_where=sa.text("superseded_by IS NOT NULL"),
+        )
+
+    # Re-inspect after potential DDL changes above (link_type column, index).
+    inspector = _inspector()
+    checkpoint_fks = inspector.get_foreign_keys("checkpoint_metadata")
+    fk_signatures = {
+        (
+            tuple(fk.get("constrained_columns") or []),
+            fk.get("referred_table"),
+            tuple(fk.get("referred_columns") or []),
+        )
+        for fk in checkpoint_fks
+    }
+
+    needs_decision_fk = (
+        ("decision_id",),
+        "decisions",
+        ("decision_id",),
+    ) not in fk_signatures
+    needs_resource_fk = (
+        ("resource_id",),
+        "resources",
+        ("resource_id",),
+    ) not in fk_signatures
+
+    if needs_decision_fk or needs_resource_fk:
+        # Nullify orphan references before creating FK constraints so the
+        # migration does not fail on existing data with dangling IDs.
+        # Uses NOT EXISTS instead of NOT IN for better query-plan
+        # performance on large checkpoint_metadata tables.
+        op.execute(
+            sa.text(
+                """
+                UPDATE checkpoint_metadata
+                SET decision_id = NULL
+                WHERE decision_id IS NOT NULL
+                  AND NOT EXISTS (
+                      SELECT 1 FROM decisions d
+                      WHERE d.decision_id = checkpoint_metadata.decision_id
+                  )
+                """
+            )
+        )
+        op.execute(
+            sa.text(
+                """
+                UPDATE checkpoint_metadata
+                SET resource_id = NULL
+                WHERE resource_id IS NOT NULL
+                  AND NOT EXISTS (
+                      SELECT 1 FROM resources r
+                      WHERE r.resource_id = checkpoint_metadata.resource_id
+                  )
+                """
+            )
+        )
+
+        # NOTE: Spec DDL (lines 45569, 45571) uses bare REFERENCES without
+        # an ON DELETE clause (default: NO ACTION / RESTRICT).  We use
+        # ondelete="SET NULL" so that DecisionRepository.delete() and
+        # ResourceRepository.delete() do not raise IntegrityError when a
+        # parent decision or resource is removed while checkpoint rows
+        # still reference it.
+        with op.batch_alter_table("checkpoint_metadata") as batch_op:
+            if needs_decision_fk:
+                batch_op.create_foreign_key(
+                    "fk_checkpoint_metadata_decision",
+                    "decisions",
+                    ["decision_id"],
+                    ["decision_id"],
+                    ondelete="SET NULL",
+                )
+            if needs_resource_fk:
+                batch_op.create_foreign_key(
+                    "fk_checkpoint_metadata_resource",
+                    "resources",
+                    ["resource_id"],
+                    ["resource_id"],
+                    ondelete="SET NULL",
+                )
+
+    if op.get_bind().dialect.name == "sqlite":
+        _create_sqlite_checkpoint_fk_triggers()
+
+
+def downgrade() -> None:
+    """Revert spec-parity schema updates."""
+    inspector = _inspector()
+
+    if op.get_bind().dialect.name == "sqlite":
+        _drop_sqlite_checkpoint_fk_triggers()
+
+    decision_indexes = {index["name"] for index in inspector.get_indexes("decisions")}
+    if "idx_decisions_superseded" in decision_indexes:
+        op.drop_index("idx_decisions_superseded", table_name="decisions")
+
+    checkpoint_fks = inspector.get_foreign_keys("checkpoint_metadata")
+    fk_names = {fk.get("name") for fk in checkpoint_fks}
+    if (
+        "fk_checkpoint_metadata_decision" in fk_names
+        or "fk_checkpoint_metadata_resource" in fk_names
+    ):
+        with op.batch_alter_table("checkpoint_metadata") as batch_op:
+            if "fk_checkpoint_metadata_decision" in fk_names:
+                batch_op.drop_constraint(
+                    "fk_checkpoint_metadata_decision",
+                    type_="foreignkey",
+                )
+            if "fk_checkpoint_metadata_resource" in fk_names:
+                batch_op.drop_constraint(
+                    "fk_checkpoint_metadata_resource",
+                    type_="foreignkey",
+                )
+
+    resource_link_columns = {
+        column["name"] for column in inspector.get_columns("resource_links")
+    }
+    if "link_type" in resource_link_columns:
+        check_constraints = inspector.get_check_constraints("resource_links")
+        has_link_type_ck = any(
+            ck.get("name") == "ck_resource_links_link_type" for ck in check_constraints
+        )
+        with op.batch_alter_table("resource_links") as batch_op:
+            if has_link_type_ck:
+                batch_op.drop_constraint("ck_resource_links_link_type", type_="check")
+            batch_op.drop_column("link_type")

src/cleveragents/infrastructure/database/migrations/versions/m8_002_merge_profile_rename_and_corrections.py

@@ -1,13 +1,14 @@
-"""Merge profile-rename, correction-attempts, and plans-schema heads.
+"""Merge profile-rename, correction-attempts, plans-schema, and schema-parity heads.
 
-m5_001_rename_profile_fields, m8_001_correction_attempts, and
-m8_001_align_plans_schema all branched from m4_003_plan_env_columns,
-creating three Alembic heads.  This no-op merge migration resolves
-them into a single head.
+m5_001_rename_profile_fields, m8_001_correction_attempts,
+m8_001_align_plans_schema, and m4_004_schema_parity_resource_decision_checkpoint
+all branched from m4_003_plan_env_columns, creating four Alembic heads.
+This no-op merge migration resolves them into a single head.
 
 Revision ID: m8_002_merge_profile_rename_and_corrections
 Revises: m5_001_rename_profile_fields, m8_001_correction_attempts,
-m8_001_align_plans_schema
+         m8_001_align_plans_schema,
+         m4_004_schema_parity_resource_decision_checkpoint
 Create Date: 2026-03-29 00:00:00
 
 """
@@ -20,6 +21,7 @@ down_revision: str | Sequence[str] | None = (
     "m5_001_rename_profile_fields",
     "m8_001_correction_attempts",
     "m8_001_align_plans_schema",
+    "m4_004_schema_parity_resource_decision_checkpoint",
 )
 branch_labels: str | Sequence[str] | None = None
 depends_on: str | Sequence[str] | None = None

src/cleveragents/infrastructure/database/models.py

@@ -1736,6 +1736,18 @@ class ResourceLinkModel(Base):  # type: ignore[misc]
         primary_key=True,
     )
 
+    # Link type metadata aligned with spec DDL (line 45549).
+    # NOTE: Spec DDL defines link_type as nullable (TEXT DEFAULT 'contains'
+    # without NOT NULL).  We enforce NOT NULL because a NULL link type is
+    # semantically meaningless and the sibling resource_edges table also
+    # requires a non-NULL link_type.
+    link_type = Column(
+        Text,
+        nullable=False,
+        default="contains",
+        server_default=text("'contains'"),
+    )
+
     # Timestamp (ISO-8601 string)
     created_at = Column(String(30), nullable=False)
 
@@ -1744,6 +1756,14 @@ class ResourceLinkModel(Base):  # type: ignore[misc]
             "parent_id != child_id",
             name="ck_resource_links_no_self_loop",
         ),
+        # NOTE: Spec DDL (line 45549) does not declare a CHECK constraint on
+        # link_type (the allowed values appear only in a comment).  We add this
+        # CHECK to match the sibling ck_resource_edges_link_type constraint and
+        # prevent invalid link types at the database level.
+        CheckConstraint(
+            "link_type IN ('contains', 'references', 'derived_from')",
+            name="ck_resource_links_link_type",
+        ),
         Index("ix_resource_links_child", "child_id"),
         Index("ix_resource_links_parent", "parent_id"),
     )
@@ -2791,6 +2811,12 @@ class DecisionModel(Base):  # type: ignore[misc]
         Index("ix_decisions_type", "decision_type"),
         Index("ix_decisions_created", "created_at"),
         Index("ix_decisions_plan_seq", "plan_id", "sequence_number"),
+        Index(
+            "idx_decisions_superseded",
+            "superseded_by",
+            postgresql_where=text("superseded_by IS NOT NULL"),
+            sqlite_where=text("superseded_by IS NOT NULL"),
+        ),
     )
 
     # --- Domain conversion ---
@@ -3001,14 +3027,27 @@ class CheckpointModel(Base):  # type: ignore[misc]
         nullable=False,
     )
 
-    # Optional FK to decisions (decision-aligned checkpoints)
-    decision_id = Column(String(26), nullable=True)
+    # Optional FK to decisions (decision-aligned checkpoints).
+    # NOTE: Spec DDL (line 45569) uses bare REFERENCES without ON DELETE.
+    # We use ondelete="SET NULL" so that deleting a decision does not raise
+    # IntegrityError when checkpoint rows still reference it.
+    decision_id = Column(
+        String(26),
+        ForeignKey("decisions.decision_id", ondelete="SET NULL"),
+        nullable=True,
+    )
 
     # Checkpoint type: pre_write, post_step, or manual
     checkpoint_type = Column(Text, nullable=False, default="manual")
 
-    # Optional resource association
-    resource_id = Column(String(26), nullable=True)
+    # Optional resource association.
+    # NOTE: Same ondelete="SET NULL" rationale as decision_id above
+    # (spec line 45571 uses bare REFERENCES).
+    resource_id = Column(
+        String(26),
+        ForeignKey("resources.resource_id", ondelete="SET NULL"),
+        nullable=True,
+    )
 
     # Sandbox reference (e.g. git commit hash)
     sandbox_ref = Column(Text, nullable=False)