feat(db): mesurer la derive du schema, au lieu de la supposer

DB-001. create_all() cree les tables manquantes et ne fait jamais d ALTER.
Une colonne ajoutee a un modele il y a six mois est donc absente de toute
base qui possedait deja la table, et rien ne le dit : l application demarre,
et la premiere requete qui touche cette colonne echoue a l execution. C est
la raison d etre de migrations/add_tryout_coaches.py, ecrit a la main pour
rattraper un cas. Personne ne sait combien il y en a d autres.

app/supporting_scripts/schema_report.py compare le catalogue d une base
vivante aux modeles : tables, colonnes, types, nullabilite, contraintes
d unicite, cles etrangeres, index. En **lecture seule** — il ouvre une
connexion, lit, imprime, sort. Aucun DDL, aucun DML.

Les constats sont classes par ce qu ils coutent, pas par ce qu ils sont :

- BLOCKING : les modeles l attendent, la base ne l a pas. C est la derive ;
- RISK : la base l a, aucun modele ne le decrit. Inoffensif tant que rien
  ne bouge — et **un alembic --autogenerate proposera de le supprimer**,
  avec ses donnees. C est la classe qu on lit en entier ;
- DIFFERENCE : types, nullabilite, contraintes. Chacune demande un humain.

Les types sont compares apres compilation vers le meme dialecte : opposer
String(200) a VARCHAR(200) en chaines aurait signale chaque colonne comme
differente, et un rapport qui crie partout ne se lit plus.

--check-seed-accounts repond a la question de SEC-003 a laquelle le depot
ne peut pas repondre : le compte admin/password seme par clear_db.py
existe-t-il encore, et son mot de passe est-il toujours celui-la.

13 tests le pilotent contre des bases SQLite fabriquees pour diverger d une
facon connue. Le cas qui compte le plus est la base propre : un rapport qui
crie sur une base saine ne sera pas lu, et un rapport qui dit « aucun
ecart » sur une base derivee est pire que pas de rapport — c est un feu
vert pour laisser autogenerate ecrire la difference en DROP.

docs/database-schema.md donne la suite, etape par etape, avec le piege de
DB-002 en toutes lettres : la migration initiale doit decrire la base telle
qu elle est, pas telle que les modeles la decrivent. Generer depuis les
modeles puis estampiller revient a declarer que la derive n existe pas.

Alembic n est pas ajoute aux dependances : rien ne l utilise encore, et une
dependance que rien n utilise est exactement ce que l audit reprochait
ailleurs. Le document dit a quelle etape l ajouter.

530 tests.

app/supporting_scripts/schema_report.py

@@ -0,0 +1,384 @@
+"""Compare a live database against the models (DB-001).
+
+Why this exists
+---------------
+`db.create_all()` creates missing tables and never ALTERs an existing one. A
+column added to a model months ago is therefore simply absent from any
+database that already had the table, and nothing says so: the application
+starts, and the first query touching that column fails at runtime. The audit
+called the accumulated result "les dérives" and could not measure it, because
+measuring it needs the production database.
+
+This script measures it. It is **read-only** — it opens a connection, reads
+the catalogue, prints a report and exits. It issues no DDL and no DML, and
+takes no locks beyond what reading `information_schema` takes.
+
+It is the prerequisite for everything in the DB wave: `DB-002` asks for an
+initial Alembic migration describing the **real** schema rather than the
+models', and this is what tells you what the real schema is.
+
+Usage
+-----
+    # Against whatever DATABASE_URL points at
+    python app/supporting_scripts/schema_report.py
+
+    # Against a restored copy, which is the safe way to do it first
+    python app/supporting_scripts/schema_report.py \
+        --url postgresql://user:pass@host:5432/restored_copy
+
+    # Also look for the seeded admin/password account (SEC-003)
+    python app/supporting_scripts/schema_report.py --check-seed-accounts
+
+Exit codes
+----------
+    0  the live schema matches the models
+    1  drift found — the report says what
+    2  could not connect or read the catalogue
+
+Reading the output
+------------------
+Findings are grouped by what they cost you:
+
+  BLOCKING   the application will fail at runtime — a table or column the
+             models use and the database does not have.
+  RISK       the database has something the models do not describe. Harmless
+             to the running application, but an Alembic autogenerate would
+             propose to DROP it, which is how a corrective migration deletes
+             a column somebody still needed.
+  DIFFERENCE type, nullability, default or constraint disagreements. Each one
+             needs a human: some are dialect spelling, some are real.
+"""
+
+import argparse
+import os
+import sys
+
+# Importable as a script from the project root, like the other supporting
+# scripts: `python app/supporting_scripts/schema_report.py`.
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))
+
+from sqlalchemy import create_engine, inspect  # noqa: E402
+from sqlalchemy.exc import SQLAlchemyError  # noqa: E402
+
+BLOCKING = 'BLOCKING'
+RISK = 'RISK'
+DIFFERENCE = 'DIFFERENCE'
+
+
+class Finding:
+    """One disagreement between the models and the live database."""
+
+    def __init__(self, severity, table, detail, consequence=''):
+        self.severity = severity
+        self.table = table
+        self.detail = detail
+        self.consequence = consequence
+
+    def __str__(self):
+        line = f'  [{self.severity:10}] {self.table}: {self.detail}'
+        if self.consequence:
+            line += f'\n               → {self.consequence}'
+        return line
+
+    def __repr__(self):  # pragma: no cover — debugging aid
+        return f'<Finding {self.severity} {self.table} {self.detail}>'
+
+
+def model_metadata():
+    """The schema the models describe.
+
+    Imports app.models for its side effect: importing the modules is what
+    registers every table on the shared metadata.
+    """
+    from app.extensions import db
+    from app.models import User  # noqa: F401 — registers the whole model package
+
+    return db.metadata
+
+
+def _type_of(column_type, dialect):
+    """A type as this dialect spells it, so the two sides are comparable.
+
+    Comparing `String(200)` with `VARCHAR(200)` as strings would report every
+    column as different. Compiling both against the same dialect makes the
+    comparison mean something.
+    """
+    try:
+        return column_type.compile(dialect=dialect)
+    except Exception:  # noqa: BLE001 — an uncompilable type is still reportable
+        return str(column_type)
+
+
+def compare_tables(metadata, inspector):
+    """Tables the models expect against tables the database has."""
+    findings = []
+    model_tables = set(metadata.tables)
+    live_tables = set(inspector.get_table_names())
+
+    for name in sorted(model_tables - live_tables):
+        findings.append(
+            Finding(
+                BLOCKING,
+                name,
+                'table is missing from the database',
+                'every query against this model fails. create_all() would '
+                'create it — which is why the absence can survive unnoticed '
+                'on a machine where AUTO_CREATE_TABLES is on.',
+            )
+        )
+
+    for name in sorted(live_tables - model_tables):
+        findings.append(
+            Finding(
+                RISK,
+                name,
+                'table exists in the database and in no model',
+                'an Alembic autogenerate would propose to DROP it. Decide '
+                'before running one: it may be a leftover, or it may be the '
+                'only copy of something.',
+            )
+        )
+
+    return findings, sorted(model_tables & live_tables)
+
+
+def compare_columns(metadata, inspector, table_name, dialect):
+    """Column-by-column, for one table."""
+    findings = []
+    model_columns = {c.name: c for c in metadata.tables[table_name].columns}
+    live_columns = {c['name']: c for c in inspector.get_columns(table_name)}
+
+    for name in sorted(set(model_columns) - set(live_columns)):
+        column = model_columns[name]
+        findings.append(
+            Finding(
+                BLOCKING,
+                table_name,
+                f'column "{name}" is in the model and not in the database',
+                'this is exactly what create_all() cannot fix: it never '
+                'ALTERs. Any query selecting or writing this column fails.'
+                + (
+                    ''
+                    if column.nullable
+                    else ' The column is NOT NULL, so the '
+                    'corrective migration needs a default or a backfill.'
+                ),
+            )
+        )
+
+    for name in sorted(set(live_columns) - set(model_columns)):
+        findings.append(
+            Finding(
+                RISK,
+                table_name,
+                f'column "{name}" is in the database and not in any model',
+                'an autogenerated migration would propose to DROP it, taking '
+                'its data. Check whether something outside the application '
+                'reads it before agreeing.',
+            )
+        )
+
+    for name in sorted(set(model_columns) & set(live_columns)):
+        model_column, live_column = model_columns[name], live_columns[name]
+
+        model_type = _type_of(model_column.type, dialect)
+        live_type = _type_of(live_column['type'], dialect)
+        if model_type != live_type:
+            findings.append(
+                Finding(
+                    DIFFERENCE,
+                    table_name,
+                    f'column "{name}" type: model says {model_type}, database says {live_type}',
+                    'a narrower column in the database silently truncates or '
+                    'rejects; a wider one is usually harmless.',
+                )
+            )
+
+        if bool(model_column.nullable) != bool(live_column.get('nullable', True)):
+            findings.append(
+                Finding(
+                    DIFFERENCE,
+                    table_name,
+                    f'column "{name}" nullability: model says '
+                    f'{"NULL" if model_column.nullable else "NOT NULL"}, database says '
+                    f'{"NULL" if live_column.get("nullable", True) else "NOT NULL"}',
+                    'a NOT NULL the database does not enforce is a constraint '
+                    'the application only believes it has.',
+                )
+            )
+
+    return findings
+
+
+def compare_constraints(metadata, inspector, table_name):
+    """Unique constraints, indexes and foreign keys.
+
+    Named constraints are compared by the columns they cover rather than by
+    name: the same rule declared under two names is the same rule, and
+    reporting it as a difference would bury the ones that matter.
+    """
+    findings = []
+    table = metadata.tables[table_name]
+
+    def column_sets(entries, key):
+        return {tuple(sorted(entry[key] or [])) for entry in entries}
+
+    model_unique = {
+        tuple(sorted(c.name for c in constraint.columns))
+        for constraint in table.constraints
+        if constraint.__class__.__name__ == 'UniqueConstraint'
+    }
+    live_unique = column_sets(inspector.get_unique_constraints(table_name), 'column_names')
+    for columns in sorted(model_unique - live_unique):
+        findings.append(
+            Finding(
+                DIFFERENCE,
+                table_name,
+                f'unique constraint on {list(columns)} is declared and absent from the database',
+                'the application believes duplicates are impossible here. '
+                'They are not, and two concurrent requests will prove it.',
+            )
+        )
+
+    model_fks = {
+        tuple(sorted(fk.parent.name for fk in constraint.elements))
+        for constraint in table.foreign_key_constraints
+    }
+    live_fks = column_sets(inspector.get_foreign_keys(table_name), 'constrained_columns')
+    for columns in sorted(model_fks - live_fks):
+        findings.append(
+            Finding(
+                DIFFERENCE,
+                table_name,
+                f'foreign key on {list(columns)} is declared and absent from the database',
+                'orphan rows are possible, and ON DELETE behaviour is not '
+                'being enforced by the database at all.',
+            )
+        )
+
+    model_indexes = {tuple(sorted(c.name for c in index.columns)) for index in table.indexes}
+    live_indexes = column_sets(inspector.get_indexes(table_name), 'column_names')
+    for columns in sorted(model_indexes - live_indexes):
+        findings.append(
+            Finding(
+                DIFFERENCE,
+                table_name,
+                f'index on {list(columns)} is declared and absent from the database',
+                'correctness is unaffected; the queries that rely on it are '
+                'doing sequential scans.',
+            )
+        )
+
+    return findings
+
+
+def collect_findings(engine):
+    """Every disagreement between the models and this database."""
+    metadata = model_metadata()
+    inspector = inspect(engine)
+
+    findings, shared_tables = compare_tables(metadata, inspector)
+    for table_name in shared_tables:
+        findings.extend(compare_columns(metadata, inspector, table_name, engine.dialect))
+        findings.extend(compare_constraints(metadata, inspector, table_name))
+    return findings
+
+
+def find_seed_accounts(engine):
+    """Accounts matching the credentials clear_db.py used to seed (SEC-003).
+
+    The script was removed from the deployment, but it had already been run:
+    the audit could not tell whether an `admin` account with the password
+    `password` still exists in production, and that question cannot be
+    answered from the repository.
+
+    Returns:
+        list[tuple]: (username, role, whether the known password matches).
+    """
+    from sqlalchemy import text
+
+    from app.extensions import check_password
+
+    with engine.connect() as connection:
+        rows = connection.execute(
+            text('SELECT username, role, password_hash FROM users WHERE username = :name'),
+            {'name': 'admin'},
+        ).fetchall()
+
+    results = []
+    for username, role, password_hash in rows:
+        try:
+            matches = check_password(password_hash, 'password')
+        except Exception:  # noqa: BLE001 — an unreadable hash is not a match
+            matches = False
+        results.append((username, role, matches))
+    return results
+
+
+def main(argv=None):
+    parser = argparse.ArgumentParser(description=__doc__.split('\n')[0])
+    parser.add_argument(
+        '--url',
+        default=os.getenv('DATABASE_URL'),
+        help='Database URL. Defaults to DATABASE_URL. Point it at a restored copy the first time.',
+    )
+    parser.add_argument(
+        '--check-seed-accounts',
+        action='store_true',
+        help='Also look for the admin/password account seeded by clear_db.py (SEC-003).',
+    )
+    args = parser.parse_args(argv)
+
+    if not args.url:
+        print('No database URL. Pass --url or set DATABASE_URL.', file=sys.stderr)
+        return 2
+
+    from app.app import normalise_database_url
+
+    try:
+        engine = create_engine(normalise_database_url(args.url))
+        findings = collect_findings(engine)
+    except SQLAlchemyError as exc:
+        print(f'Could not read the schema: {exc}', file=sys.stderr)
+        return 2
+
+    print('=' * 78)
+    print('Schema report — models vs live database (DB-001)')
+    print('=' * 78)
+
+    if not findings:
+        print('\nNo drift. The live schema matches the models.')
+    for severity in (BLOCKING, RISK, DIFFERENCE):
+        group = [f for f in findings if f.severity == severity]
+        if not group:
+            continue
+        print(f'\n{severity} — {len(group)} finding(s)')
+        for finding in group:
+            print(finding)
+
+    if args.check_seed_accounts:
+        print('\n' + '=' * 78)
+        print('Seeded accounts (SEC-003)')
+        print('=' * 78)
+        try:
+            accounts = find_seed_accounts(engine)
+        except SQLAlchemyError as exc:
+            print(f'Could not check: {exc}')
+        else:
+            if not accounts:
+                print('No account named "admin".')
+            for username, role, matches in accounts:
+                verdict = (
+                    'PASSWORD IS STILL "password" — change it now'
+                    if matches
+                    else 'password has been changed'
+                )
+                print(f'  {username} ({role}): {verdict}')
+
+    blocking = sum(1 for f in findings if f.severity == BLOCKING)
+    print(f'\n{len(findings)} finding(s), {blocking} blocking.')
+    return 1 if findings else 0
+
+
+if __name__ == '__main__':  # pragma: no cover
+    sys.exit(main())

docs/database-schema.md

@@ -0,0 +1,121 @@
+# Le schéma réel, et comment sortir de `create_all()`
+
+> **État au 2026-08-11** : l'outil de relevé existe et est testé. Le relevé
+> lui-même n'a pas été exécuté — il demande un accès à la base de production,
+> qui ne peut pas venir du dépôt. Tout ce qui suit attend cette exécution.
+
+## Pourquoi c'est le nœud
+
+`db.create_all()` crée les tables manquantes et **ne fait jamais d'`ALTER`**.
+
+Une colonne ajoutée à un modèle il y a six mois est donc absente de toute base
+qui possédait déjà la table, et rien ne le dit : l'application démarre
+normalement, et la première requête qui touche cette colonne échoue à
+l'exécution. C'est la raison d'être de `migrations/add_tryout_coaches.py`, un
+script écrit à la main pour rattraper un cas.
+
+Personne ne sait combien il y en a d'autres. C'est ce que `DB-001` mesure, et
+c'est pourquoi **huit tâches en dépendent** : `DB-002` à `DB-009`, plus
+`ARCH-001` (fusion coach/équipe) et `SEC-012` (identité Discord avec
+`unique=True`).
+
+## Étape 1 — sauvegarder, et vérifier la sauvegarde
+
+**Rien de ce qui suit ne se fait avant qu'une sauvegarde ait été restaurée
+avec succès.** Pas « prise » : *restaurée*. Une sauvegarde qu'on n'a jamais
+restaurée est une hypothèse.
+
+Procédure dans `docs/database-restore.md`. La copie restaurée sert aussi de
+terrain pour les étapes 2 et 3.
+
+## Étape 2 — relever l'écart
+
+```bash
+# D'abord sur la copie restaurée, jamais directement sur la production
+python app/supporting_scripts/schema_report.py \
+    --url postgresql://user:pass@host:5432/copie_restauree
+```
+
+L'outil est **en lecture seule** : il ouvre une connexion, lit le catalogue,
+imprime et sort. Aucun DDL, aucun DML.
+
+Codes de sortie : `0` aucun écart, `1` écart trouvé, `2` connexion impossible.
+
+Le rapport classe ce qu'il trouve par ce que ça coûte :
+
+| Classe | Ce que c'est | Ce que ça coûte |
+|---|---|---|
+| `BLOCKING` | Les modèles l'attendent, la base ne l'a pas | L'application échoue à l'exécution. C'est la dérive de `create_all()` |
+| `RISK` | La base l'a, aucun modèle ne le décrit | Inoffensif tant que rien ne bouge. **Un `alembic --autogenerate` proposera de le supprimer**, avec ses données |
+| `DIFFERENCE` | Types, nullabilité, contraintes qui divergent | Chacune demande un humain : certaines sont de l'orthographe de dialecte, d'autres sont réelles |
+
+**La classe `RISK` est celle qu'on lit en entier.** C'est par là qu'une
+migration corrective détruit une colonne dont quelqu'un se servait encore.
+
+Pendant qu'on y est, la question que l'audit posait et à laquelle le dépôt ne
+peut pas répondre — le compte `admin`/`password` semé par `clear_db.py`
+existe-t-il encore ? (`SEC-003`) :
+
+```bash
+python app/supporting_scripts/schema_report.py --check-seed-accounts
+```
+
+## Étape 3 — Alembic, décrivant le schéma **réel** (`DB-002`)
+
+Le piège de cette étape tient en une phrase : **la migration initiale doit
+décrire la base telle qu'elle est, pas telle que les modèles la décrivent.**
+
+Générer la migration initiale depuis les modèles puis estampiller la
+production revient à déclarer que la dérive n'existe pas. Elle reste là,
+invisible, et la première migration suivante s'appuie sur un état faux.
+
+1. Ajouter `alembic` à `requirements.txt` — et seulement à ce moment : une
+   dépendance que rien n'utilise est exactement ce que l'audit reprochait
+   ailleurs.
+2. `alembic init migrations/alembic`, en pointant `sqlalchemy.url` sur
+   `DATABASE_URL` plutôt qu'en le codant en dur.
+3. Générer la révision initiale **contre la copie restaurée** :
+   `alembic revision --autogenerate -m "schéma existant"`.
+4. **Relire la révision ligne par ligne** contre le rapport de l'étape 2. Tout
+   `op.drop_*` correspond à une ligne `RISK` : ou bien on l'assume, ou bien on
+   le retire de la migration.
+5. Estampiller : `alembic stamp head`. La révision initiale ne s'exécute
+   jamais ; elle décrit le point de départ.
+
+## Étape 4 — la migration corrective (`DB-003`)
+
+Une seconde révision qui rattrape les écarts relevés. À exécuter d'abord sur
+la copie restaurée, et à vérifier avec :
+
+```bash
+alembic upgrade head && alembic downgrade -1 && alembic upgrade head
+python app/supporting_scripts/schema_report.py --url <copie>   # doit sortir 0
+```
+
+Le critère d'acceptation est celui-là : le relevé ne trouve plus rien.
+
+Attention aux colonnes `BLOCKING` déclarées `NOT NULL` : les ajouter à une
+table peuplée échoue sans valeur par défaut ni remplissage. Le rapport le
+signale dans la conséquence.
+
+## Étape 5 — ce que la migration débloque
+
+Dans cet ordre, parce qu'ils dépendent tous de `DB-002` :
+
+| Tâche | Ce qu'elle fait | Pourquoi ça attendait |
+|---|---|---|
+| `DB-004` | Retirer `create_all()` de `create_app()` | Tant qu'il est là, deux mécanismes décrivent le schéma |
+| `DB-005` | Cascades de suppression au niveau base | Les cascades ORM sont en place ; PostgreSQL ne les connaît pas |
+| `DB-006` | Unicité sur `TryoutRegistration(tryout_id, player_id)` | Le plafond d'inscriptions est aujourd'hui un `count()` suivi d'un `add()` : deux requêtes simultanées passent toutes les deux |
+| `DB-007` | Index, `CheckConstraint` sur les statuts, `server_default` | — |
+| `DB-008` | Trancher `attendance_confirmed` côté tryout | `discord_bot.py` écrit un attribut fantôme ; aujourd'hui journalisé en avertissement |
+| `DB-009` | Horodatages avec fuseau | `datetime.utcnow` partout, déprécié en 3.12 |
+| `ARCH-001` | Fusionner coach/équipe sur la relation m2m | Migration de données ; `app/permissions.py` rend la duplication inoffensive **en lecture** seulement, l'écriture crée toujours les deux |
+| `SEC-012` | Identité Discord côté serveur, `unique=True` | La colonne doit être unique, donc dédoublonnée d'abord |
+
+## Ce qu'on ne fait pas
+
+Générer la migration initiale depuis les modèles « pour avancer en
+attendant ». Ça produit un dépôt qui a l'air d'avoir des migrations, une
+production estampillée sur un état qu'elle n'a pas, et la dérive préservée
+sous une couche de plus.

tests/test_schema_report.py

@@ -0,0 +1,259 @@
+"""The schema drift report actually finds drift (DB-001).
+
+`db.create_all()` creates missing tables and never ALTERs one. Every column
+added to a model since a table was first created is therefore absent from
+any database that already had it, and the application says nothing until a
+query touches it. Measuring that needs the production database — which is
+why the audit could only name the problem.
+
+These tests drive the report against SQLite databases built to disagree with
+the models in a specific way, so that the tool's answer can be checked
+against a known truth. A report that returns "no drift" on a drifted
+database is worse than no report: it is a green light to run
+`alembic revision --autogenerate` and let it write the difference as DROPs.
+
+The tool is read-only; nothing here asserts otherwise, because nothing here
+gives it the chance to be anything else.
+"""
+
+import sqlite3
+
+import pytest
+from sqlalchemy import create_engine
+
+from app.supporting_scripts.schema_report import (
+    BLOCKING,
+    DIFFERENCE,
+    RISK,
+    collect_findings,
+    main,
+)
+
+
+@pytest.fixture
+def live_db(tmp_path, app):
+    """A database created from the models, then tampered with.
+
+    Built through the application so it starts out identical to the models —
+    the baseline the report has to call clean before its disagreements mean
+    anything.
+    """
+    path = tmp_path / 'live.sqlite'
+    engine = create_engine(f'sqlite:///{path}')
+
+    from app.extensions import db
+
+    with app.app_context():
+        db.metadata.create_all(engine)
+
+    def tamper(*statements):
+        connection = sqlite3.connect(path)
+        for statement in statements:
+            connection.execute(statement)
+        connection.commit()
+        connection.close()
+        return create_engine(f'sqlite:///{path}')
+
+    return engine, tamper
+
+
+def _by_severity(findings, severity):
+    return [f for f in findings if f.severity == severity]
+
+
+class TestCleanDatabase:
+    def test_a_database_built_from_the_models_reports_no_drift(self, live_db, app):
+        engine, _tamper = live_db
+
+        with app.app_context():
+            findings = collect_findings(engine)
+
+        assert findings == [], f'false positives make the report unusable: {findings}'
+
+
+class TestMissingThings:
+    """The BLOCKING class: the models expect it, the database has not got it."""
+
+    def test_a_dropped_table_is_blocking(self, live_db, app):
+        engine, tamper = live_db
+        engine = tamper('DROP TABLE evaluations')
+
+        with app.app_context():
+            findings = collect_findings(engine)
+
+        blocking = _by_severity(findings, BLOCKING)
+        assert any(f.table == 'evaluations' for f in blocking)
+
+    def test_a_column_the_model_added_later_is_blocking(self, live_db, app):
+        """The exact shape of the create_all() drift: the table exists, so
+        create_all() leaves it alone for ever."""
+        engine, tamper = live_db
+        engine = tamper(
+            'ALTER TABLE tryouts RENAME TO tryouts_old',
+            'CREATE TABLE tryouts (id INTEGER PRIMARY KEY, title VARCHAR(200) NOT NULL)',
+        )
+
+        with app.app_context():
+            findings = collect_findings(engine)
+
+        missing = [f for f in _by_severity(findings, BLOCKING) if f.table == 'tryouts']
+        assert any('"game"' in f.detail for f in missing)
+        assert any('"date"' in f.detail for f in missing)
+
+    def test_a_missing_not_null_column_says_the_migration_needs_a_default(self, live_db, app):
+        """A corrective migration that adds a NOT NULL column to a populated
+        table fails unless it carries a default or a backfill. The report has
+        to say so, because that is where the migration breaks."""
+        engine, tamper = live_db
+        engine = tamper(
+            'ALTER TABLE tryouts RENAME TO tryouts_old',
+            'CREATE TABLE tryouts (id INTEGER PRIMARY KEY, title VARCHAR(200) NOT NULL)',
+        )
+
+        with app.app_context():
+            findings = collect_findings(engine)
+
+        game = next(f for f in findings if '"game"' in f.detail)
+        assert 'NOT NULL' in game.consequence
+        assert 'backfill' in game.consequence
+
+
+class TestExtraThings:
+    """The RISK class: the database has it, no model describes it.
+
+    Harmless while running, dangerous the moment somebody autogenerates a
+    migration — which will propose to DROP it.
+    """
+
+    def test_an_unknown_table_is_reported_as_a_risk(self, live_db, app):
+        engine, tamper = live_db
+        engine = tamper('CREATE TABLE legacy_import (id INTEGER PRIMARY KEY)')
+
+        with app.app_context():
+            findings = collect_findings(engine)
+
+        risks = _by_severity(findings, RISK)
+        assert any(f.table == 'legacy_import' for f in risks)
+        assert any('DROP' in f.consequence for f in risks)
+
+    def test_an_unknown_column_is_reported_as_a_risk(self, live_db, app):
+        engine, tamper = live_db
+        engine = tamper('ALTER TABLE tryouts ADD COLUMN old_notes TEXT')
+
+        with app.app_context():
+            findings = collect_findings(engine)
+
+        assert any(f.severity == RISK and 'old_notes' in f.detail for f in findings)
+
+
+class TestDifferences:
+    def test_a_nullability_disagreement_is_reported(self, live_db, app):
+        """A NOT NULL the database does not enforce is a constraint the
+        application only believes it has."""
+        engine, tamper = live_db
+        engine = tamper(
+            'ALTER TABLE tryouts RENAME TO tryouts_old',
+            'CREATE TABLE tryouts ('
+            ' id INTEGER PRIMARY KEY, title VARCHAR(200), game VARCHAR(50) NOT NULL,'
+            ' date DATE NOT NULL, created_by INTEGER NOT NULL)',
+        )
+
+        with app.app_context():
+            findings = collect_findings(engine)
+
+        assert any(
+            f.severity == DIFFERENCE and 'nullability' in f.detail and '"title"' in f.detail
+            for f in findings
+        )
+
+    def test_a_type_disagreement_is_reported(self, live_db, app):
+        engine, tamper = live_db
+        engine = tamper(
+            'ALTER TABLE tryouts RENAME TO tryouts_old',
+            'CREATE TABLE tryouts ('
+            ' id INTEGER PRIMARY KEY, title VARCHAR(20) NOT NULL, game VARCHAR(50) NOT NULL,'
+            ' date DATE NOT NULL, created_by INTEGER NOT NULL)',
+        )
+
+        with app.app_context():
+            findings = collect_findings(engine)
+
+        narrowed = [f for f in findings if 'type' in f.detail and '"title"' in f.detail]
+        assert narrowed, 'a column narrowed from 200 to 20 characters truncates silently'
+        assert 'VARCHAR(20)' in narrowed[0].detail
+
+
+class TestExitCodes:
+    """The report is meant to be runnable from a check, so the code matters."""
+
+    def test_a_clean_database_exits_zero(self, live_db, app, capsys):
+        engine, _tamper = live_db
+
+        with app.app_context():
+            assert main(['--url', str(engine.url)]) == 0
+
+        assert 'No drift' in capsys.readouterr().out
+
+    def test_a_drifted_database_exits_one(self, live_db, app, capsys):
+        engine, tamper = live_db
+        engine = tamper('DROP TABLE evaluations')
+
+        with app.app_context():
+            assert main(['--url', str(engine.url)]) == 1
+
+        assert 'BLOCKING' in capsys.readouterr().out
+
+    def test_no_url_at_all_exits_two_rather_than_guessing(self, monkeypatch, capsys):
+        """Defaulting to something would eventually point this at the wrong
+        database."""
+        monkeypatch.delenv('DATABASE_URL', raising=False)
+
+        assert main([]) == 2
+
+
+class TestSeedAccountCheck:
+    def test_it_finds_an_admin_whose_password_is_still_the_seeded_one(self, live_db, app, capsys):
+        """SEC-003 asked whether clear_db.py's admin/password account still
+        exists in production. That cannot be answered from the repository."""
+        engine, tamper = live_db
+
+        from app.extensions import hash_password
+
+        with app.app_context():
+            weak = hash_password('password')
+        connection = sqlite3.connect(str(engine.url).replace('sqlite:///', ''))
+        connection.execute(
+            'INSERT INTO users (username, password_hash, role, full_name, email, is_active_account)'
+            " VALUES ('admin', ?, 'admin', 'Seeded', '[email protected]', 1)",
+            (weak,),
+        )
+        connection.commit()
+        connection.close()
+
+        with app.app_context():
+            main(['--url', str(engine.url), '--check-seed-accounts'])
+
+        assert 'PASSWORD IS STILL' in capsys.readouterr().out
+
+    def test_a_changed_password_reports_as_changed(self, live_db, app, capsys):
+        engine, _tamper = live_db
+
+        from app.extensions import hash_password
+
+        with app.app_context():
+            strong = hash_password('a-real-password-9X')
+        connection = sqlite3.connect(str(engine.url).replace('sqlite:///', ''))
+        connection.execute(
+            'INSERT INTO users (username, password_hash, role, full_name, email, is_active_account)'
+            " VALUES ('admin', ?, 'admin', 'Real', '[email protected]', 1)",
+            (strong,),
+        )
+        connection.commit()
+        connection.close()
+
+        with app.app_context():
+            main(['--url', str(engine.url), '--check-seed-accounts'])
+
+        output = capsys.readouterr().out
+        assert 'password has been changed' in output
+        assert 'PASSWORD IS STILL' not in output