diff --git a/AGENTS.md b/AGENTS.md
index f0dbd8c..a86f1e9 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -74,7 +74,12 @@ See the `guide/` directory for detailed documentation:
 
 - TDD: Red -> Green -> Refactor
 - Tidy First: separate structural changes from behavioral changes
-- Use GORM queries only (no raw SQL)
+- Use GORM's model layer for queries. Raw SQL is allowed only where GORM has no
+  form for the statement — full-text operators, FTS5 virtual-table DDL and
+  writes, schema introspection its migrator cannot do, connection pragmas — and
+  only inside `internal/adapters/outbound/searchsql` and `internal/db`.
+  Identifiers come from package constants, values are always bound parameters.
+  Details and the evidence for each exemption: `guide/development.md` §Raw SQL
 - Tests: `CGO_ENABLED=1 go test -tags "fts5" ./... -count=1`
 - Integration test: `./scripts/integration-test.sh` (full Gitea + PostgreSQL + ccg Docker pipeline)
 
diff --git a/CLAUDE.md b/CLAUDE.md
index e0dadd4..f7c1baf 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -57,7 +57,7 @@ Graceful shutdown: SIGINT/SIGTERM 시 진행 중인 clone/build에 context cance
 
 - TDD: Red → Green → Refactor
 - Tidy First: 구조적 변경과 행위 변경 분리
-- GORM 쿼리만 사용 (raw SQL 금지)
+- 쿼리는 GORM 모델 계층으로 작성한다. Raw SQL은 GORM에 대응 형태가 없는 경우에만 허용한다 — 전문 검색 연산자, FTS5 가상 테이블 DDL·쓰기, migrator가 못 하는 스키마 introspection, 연결 pragma — 그리고 `internal/adapters/outbound/searchsql`와 `internal/db` 안에서만이다. 식별자는 패키지 상수에서 오고, 값은 항상 bound parameter다 (상세와 각 예외의 근거: `guide/development.md` §Raw SQL)
 - 코드 정렬: 종류별 그룹화가 아니라 "타입 + 그 타입의 생성자·메소드"를 붙여 두는 응집 관례를 따른다 (상세: `guide/development.md` §Declaration order)
 - 테스트: `CGO_ENABLED=1 go test -tags "fts5" ./... -count=1`
 - Integration test: `./scripts/integration-test.sh` (Gitea + PostgreSQL + ccg Docker 전체 파이프라인)
diff --git a/guide/development.md b/guide/development.md
index 10b82f8..ddb45f8 100644
--- a/guide/development.md
+++ b/guide/development.md
@@ -205,11 +205,56 @@ go test ./internal/adapters/inbound/cli -run TestProjectSkills -count=1
 
 - TDD: Red → Green → Refactor
 - Tidy First: Separate structural changes from behavioral changes
-- Use GORM queries only (no raw SQL)
+- Use GORM's model layer for queries; raw SQL only where GORM has no form for
+  the statement (see [Raw SQL](#raw-sql))
 - Logging: `slog`
 - CLI: `cobra` framework
 - Build flags: `CGO_ENABLED=1 -tags "fts5"`
 
+### Raw SQL
+
+Write queries through GORM's model layer — `Model`, `Where`, `FindInBatches`,
+`Migrator` — wherever GORM has a form for the statement. That is the default, and
+outside the two packages named below a raw `Raw`/`Exec` is a review stop.
+
+Raw SQL is allowed only where GORM has no form at all. That is not a matter of
+taste; each category below names something GORM's builder or migrator cannot
+express:
+
+- **Full-text operators and index maintenance** — SQLite FTS5 `MATCH` and its
+  `rank` column; PostgreSQL `to_tsvector`, `to_tsquery`, `ts_rank`, `@@`. GORM
+  has no builder form for a match operator or a rank expression.
+- **DDL on the FTS5 virtual tables** — `CREATE VIRTUAL TABLE … USING fts5`,
+  and the `DROP`/`ALTER TABLE … RENAME TO` pairs the legacy upgrade needs.
+  `AutoMigrate` does not model a virtual table.
+- **Writes into the FTS5 virtual tables** — the namespace-scoped deletes and the
+  bulk inserts. These tables have no GORM model and are not in `AutoMigrate`;
+  routing them through `Table("search_fts")` would name the table in a string
+  either way, and the bulk insert is one statement on purpose so a rebuild does
+  not pay a round trip per row.
+- **Schema introspection GORM's migrator cannot do** — `PRAGMA table_info`,
+  `sqlite_master`, `information_schema.columns`, `pg_indexes`, `pg_trigger`.
+- **Connection pragmas** — `PRAGMA journal_mode`, `PRAGMA busy_timeout`.
+
+These live in `internal/adapters/outbound/searchsql` and `internal/db` only.
+
+Two constraints hold inside an exempt statement. Table and column names come
+from package constants, never from caller input. Every value is a bound
+parameter — `Exec("DELETE FROM "+sqliteFTSTable+" WHERE namespace = ?", ns)` is
+correct; concatenating `ns` into the string is not, and no amount of exemption
+makes it correct.
+
+The introspection exemption is the one worth checking rather than trusting,
+because GORM does ship `Migrator().HasTable`, `HasColumn` and `ColumnTypes`.
+`searchsql/migrator_limits_test.go` runs all three against a real FTS5 table and
+records what happens: `HasTable` works, `ColumnTypes` fails with `invalid DDL`,
+and `HasColumn` matches the DDL text rather than the schema, so it can report
+`false` for a column that exists. `HasTable` works but returns no error, while
+every caller of `sqliteTableExists` propagates one — swallowing it would turn a
+transient failure into "the table is absent" in the upgrade path. If a GORM
+upgrade makes that test fail, the exemption should be reconsidered, not the
+test relaxed.
+
 ### Declaration order within a file
 
 Follow the standard-library convention of **cohesion over kind-grouping**: keep a
diff --git a/guide/ko/development.md b/guide/ko/development.md
index 720f123..dc08c7b 100644
--- a/guide/ko/development.md
+++ b/guide/ko/development.md
@@ -207,7 +207,48 @@ go test ./internal/adapters/inbound/cli -run TestProjectSkills -count=1
 
 - TDD: Red → Green → Refactor
 - Tidy First: 구조적 변경과 행동 변경의 분리
-- GORM 쿼리만 사용 (Raw SQL 사용 금지)
+- 쿼리는 GORM 모델 계층으로 작성. Raw SQL은 GORM에 대응 형태가 없는 경우에만 허용
+  ([Raw SQL](#raw-sql) 참고)
 - 로깅: `slog`
 - CLI: `cobra` 프레임워크
 - 빌드 플래그: `CGO_ENABLED=1 -tags "fts5"`
+
+### Raw SQL
+
+GORM에 대응 형태가 있는 문장은 모두 GORM 모델 계층(`Model`, `Where`,
+`FindInBatches`, `Migrator`)으로 작성합니다. 이것이 기본이고, 아래 두 패키지
+밖에서 나타나는 `Raw`/`Exec`는 리뷰에서 멈춰야 합니다.
+
+Raw SQL은 GORM에 대응 형태가 아예 없는 경우에만 허용합니다. 취향 문제가 아니라,
+아래 각 항목은 GORM builder나 migrator가 표현할 수 없는 것을 가리킵니다.
+
+- **전문 검색 연산자와 인덱스 관리** — SQLite FTS5의 `MATCH`와 `rank` 컬럼,
+  PostgreSQL의 `to_tsvector`, `to_tsquery`, `ts_rank`, `@@`. match 연산자나 rank
+  표현식에 해당하는 builder 형태가 GORM에 없습니다.
+- **FTS5 가상 테이블 DDL** — `CREATE VIRTUAL TABLE … USING fts5`, 그리고 legacy
+  업그레이드가 필요로 하는 `DROP` / `ALTER TABLE … RENAME TO` 쌍. `AutoMigrate`는
+  가상 테이블을 모델링하지 않습니다.
+- **FTS5 가상 테이블 쓰기** — namespace 범위 delete와 bulk insert. 이 테이블들은
+  GORM 모델이 없고 `AutoMigrate` 대상도 아니어서, `Table("search_fts")`를 거쳐도
+  테이블 이름은 결국 문자열로 남습니다. bulk insert가 한 문장인 것은 의도이며,
+  rebuild가 행마다 round trip을 내지 않게 합니다.
+- **GORM migrator가 못 하는 스키마 introspection** — `PRAGMA table_info`,
+  `sqlite_master`, `information_schema.columns`, `pg_indexes`, `pg_trigger`.
+- **연결 pragma** — `PRAGMA journal_mode`, `PRAGMA busy_timeout`.
+
+허용 범위는 `internal/adapters/outbound/searchsql`와 `internal/db` 두 패키지뿐입니다.
+
+허용된 문장 안에서도 두 제약이 유지됩니다. 테이블·컬럼 이름은 패키지 상수에서만
+오고, 호출자 입력에서 오지 않습니다. 값은 항상 bound parameter입니다 —
+`Exec("DELETE FROM "+sqliteFTSTable+" WHERE namespace = ?", ns)`는 맞고, `ns`를
+문자열에 이어 붙이는 것은 틀립니다. 어떤 예외도 이것을 맞게 만들지 않습니다.
+
+introspection 예외는 믿지 말고 확인해야 하는 항목입니다. GORM에도
+`Migrator().HasTable`, `HasColumn`, `ColumnTypes`가 있기 때문입니다.
+`searchsql/migrator_limits_test.go`가 실제 FTS5 테이블을 상대로 셋 다 실행하고
+결과를 기록합니다: `HasTable`은 동작하고, `ColumnTypes`는 `invalid DDL`로
+실패하며, `HasColumn`은 스키마가 아니라 DDL 텍스트를 매칭하므로 존재하는 컬럼에도
+`false`를 반환할 수 있습니다. `HasTable`은 동작하지만 error를 돌려주지 않고,
+`sqliteTableExists`의 모든 호출자는 error를 전파합니다. 그것을 삼키면 업그레이드
+경로에서 일시적 실패가 "테이블 없음"으로 바뀝니다. GORM 업그레이드로 그 테스트가
+깨지면 테스트를 느슨하게 만들 것이 아니라 예외 자체를 다시 판단해야 합니다.
diff --git a/internal/adapters/outbound/searchsql/backend.go b/internal/adapters/outbound/searchsql/backend.go
index 2f409d1..cbc4aa6 100644
--- a/internal/adapters/outbound/searchsql/backend.go
+++ b/internal/adapters/outbound/searchsql/backend.go
@@ -1,4 +1,11 @@
 // @index Shared search backend interface and errors for SQLite FTS5 and PostgreSQL tsvector implementations.
+//
+// The raw SQL in this package is bounded and deliberate: full-text operators,
+// DDL and writes against the FTS5 virtual tables, and the schema introspection
+// GORM's migrator cannot do. guide/development.md §Raw SQL states the rule and
+// what falls under it; migrator_limits_test.go holds the evidence for the
+// introspection part. Anything GORM's model layer can express belongs there
+// instead, here as much as anywhere else.
 package searchsql
 
 import (
diff --git a/internal/adapters/outbound/searchsql/migrator_limits_test.go b/internal/adapters/outbound/searchsql/migrator_limits_test.go
new file mode 100644
index 0000000..ca3758b
--- /dev/null
+++ b/internal/adapters/outbound/searchsql/migrator_limits_test.go
@@ -0,0 +1,100 @@
+//go:build fts5
+
+package searchsql
+
+import (
+	"testing"
+
+	"gorm.io/driver/sqlite"
+	"gorm.io/gorm"
+	"gorm.io/gorm/logger"
+)
+
+// This file is the evidence behind one line of guide/development.md §Raw SQL:
+// that the schema introspection in this package cannot go through GORM's
+// migrator. GORM does ship HasTable, HasColumn and ColumnTypes, so the exemption
+// would be a bare assertion without a test that runs all three against a real
+// FTS5 virtual table and records what they do.
+//
+// If a GORM upgrade makes one of these fail, the right response is to
+// reconsider the exemption for that call, not to relax the test.
+
+// newVirtualTableProbeDB opens an in-memory database holding the FTS5 table this
+// package creates in production, plus a second one shaped like the pre-namespace
+// legacy table.
+func newVirtualTableProbeDB(t *testing.T) *gorm.DB {
+	t.Helper()
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{Logger: logger.Discard})
+	if err != nil {
+		t.Fatalf("open sqlite: %v", err)
+	}
+	if err := createSQLiteFTSTable(db, sqliteFTSTable, true); err != nil {
+		t.Fatalf("create %s: %v", sqliteFTSTable, err)
+	}
+	if err := db.Exec(
+		"CREATE VIRTUAL TABLE legacy_probe USING fts5(node_id UNINDEXED, content, language)",
+	).Error; err != nil {
+		t.Fatalf("create legacy_probe: %v", err)
+	}
+	return db
+}
+
+// TestMigratorHasTableSeesFTS5VirtualTables records that table presence is the
+// one introspection GORM answers correctly here — and why sqliteTableExists
+// keeps its raw statement anyway: HasTable returns a bool with no error, while
+// every caller of sqliteTableExists propagates one. Swallowing that error would
+// read a transient failure as "the table is absent", which in the legacy-upgrade
+// path is the difference between stopping and rebuilding.
+func TestMigratorHasTableSeesFTS5VirtualTables(t *testing.T) {
+	db := newVirtualTableProbeDB(t)
+
+	if !db.Migrator().HasTable(sqliteFTSTable) {
+		t.Errorf("HasTable(%q) = false, want true: the virtual table was just created", sqliteFTSTable)
+	}
+	if db.Migrator().HasTable("no_such_table") {
+		t.Error("HasTable(\"no_such_table\") = true, want false")
+	}
+
+	exists, err := sqliteTableExists(db, sqliteFTSTable)
+	if err != nil || !exists {
+		t.Errorf("sqliteTableExists(%q) = (%v, %v), want (true, nil)", sqliteFTSTable, exists, err)
+	}
+}
+
+// TestMigratorColumnTypesRejectsFTS5VirtualTables records that ColumnTypes
+// cannot describe a virtual table at all: it parses the stored DDL, and
+// CREATE VIRTUAL TABLE is not a shape it can parse.
+func TestMigratorColumnTypesRejectsFTS5VirtualTables(t *testing.T) {
+	db := newVirtualTableProbeDB(t)
+
+	types, err := db.Migrator().ColumnTypes(sqliteFTSTable)
+	if err == nil {
+		t.Fatalf("ColumnTypes(%q) succeeded with %d columns, want an error", sqliteFTSTable, len(types))
+	}
+	if got := err.Error(); got != "invalid DDL" {
+		t.Errorf("ColumnTypes(%q) error = %q, want %q", sqliteFTSTable, got, "invalid DDL")
+	}
+}
+
+// TestMigratorHasColumnMisreadsFTS5VirtualTables is the reason PRAGMA table_info
+// stays. HasColumn matches patterns against the stored DDL text rather than
+// asking the schema, so whether it finds a column depends on how that column
+// happens to be spelled in the CREATE statement. `content` is a real column on
+// legacy_probe and HasColumn says it is not there.
+func TestMigratorHasColumnMisreadsFTS5VirtualTables(t *testing.T) {
+	db := newVirtualTableProbeDB(t)
+
+	present, err := sqliteColumnExists(db, "legacy_probe", "content")
+	if err != nil {
+		t.Fatalf("sqliteColumnExists: %v", err)
+	}
+	if !present {
+		t.Fatal("PRAGMA table_info does not see legacy_probe.content; the fixture is wrong, not GORM")
+	}
+
+	if db.Migrator().HasColumn("legacy_probe", "content") {
+		t.Error("HasColumn now agrees with the schema on this table. GORM may have gained real " +
+			"virtual-table introspection: reconsider the exemption in guide/development.md §Raw SQL " +
+			"rather than relaxing this test")
+	}
+}