Correções da auditoria de 24/09 (audits/fitme3070-auditoria-2026-09-24.html).
Lixeira
- Tudo o que é importante passa a soft delete (deleted_at + deleted_by):
equipa, modalidades, eventos, mensalidades, galerias, imagens, horário,
horas de funcionamento, inscrições em eventos e pedidos de inscrição.
- config/functions_trash.php: um registo por tipo; apagar, restaurar, listar
e eliminar leem todos dele. Eliminar de vez só com 30 dias na lixeira —
a condição está no WHERE do DELETE, um POST forjado não apaga nada.
- Gestão → Lixeira (views/admin/trash/index.php): lista única com quem e
quando, restauro a qualquer momento.
- Imagens de galeria: o site lista a pasta, por isso o ficheiro passa para
.trash/ (negado pelo .htaccess) e volta no restauro.
- As listagens e o site público filtram deleted_at (horário também filtra
modalidades apagadas). Migração: database/migration_20260925_trash.sql
(só acrescenta colunas; tem de ser aplicada antes deste código).
CSRF
- csrf_check() recusa qualquer pedido que não seja POST (405). Antes saía
sem verificar nada em GET, e dez ações de apagar/restaurar eram GET.
- Os dez handlers passaram a POST com token; links viraram formulários.
Autenticação e sessão
- client_ip(): X-Real-IP / último X-Forwarded-For, só vindo de proxy privado
com TRUSTED_PROXY. O throttle contava o IP do proxy (limite global).
- Throttle também por conta (hash do email), além do IP.
- Sem oráculo de tempo: login verifica contra um hash de referência quando
a conta não existe; a recuperação deixa o email na fila.
- Cookie de sessão Secure/HttpOnly/SameSite=Lax e use_strict_mode.
- Definições de sistema só para sysadmin; autocomplete=username.
Servidor e uploads
- .htaccess: nosniff, Referrer-Policy, X-Frame-Options, Permissions-Policy,
CSP (frame-ancestors, base-uri, object-src, form-action); nega parciais
_*.php, PHP soltos do platform-tools e *.md; 404 própria.
- Uploads validados pelo conteúdo (finfo/getimagesize); SVG só sem scripts,
handlers on*=, foreignObject ou hrefs externos.
Testado numa cópia da BD de produção (fitme_audit_test) com um contentor
temporário: 34/34 verificações HTTP ponta-a-ponta.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fitme3070.pt delegated all authentication to hub.andmore.pt: views/auth/*
were bounce stubs, and because this domain is outside the .andmore.pt cookie
scope the real handoff was a one-time HTTPS ticket redeem. Bring identity back
in-house — passwords live in users.password_hash, which the SSO migration
never dropped.
Removed: config/sso.php, IDENTITY_MODE, every SSO_* constant and env key,
identity_is_federated() and the cookie/ticket paths in the auth views and the
public-site bootstrap. ANDMORE_ID_URL stays — it is still the hub /notify
endpoint for outbound email/SMS, which the reset mails ride on, and has
nothing to do with authentication now.
Added, ported from the fotografia.andmore.pt sibling rather than invented:
- auth_login() with transparent password_needs_rehash upgrades and the
last_login trail; auth_register_client(); auth_set_password(); the
single-use 2h reset tokens.
- Real login / register / forgot_password / reset_password pages on a new
anonymous shell (includes/auth_header.php). includes/header.php renders
the backoffice sidebar and a logout button, which is wrong for a visitor
who is not signed in yet.
- safe_return_path(), so a crafted ?return_to= cannot bounce a freshly
authenticated user off-site.
- Per-IP throttling on login, register and reset-request. Successes count
too, so a guessed password cannot be parlayed into stuffing the rest of
the user list from the same IP.
- views/user/set_password.php: admin override for anyone whose email is
undeliverable. Prefers mailing an invite link so the backoffice never
handles the password. A plain admin cannot act on a sysadmin, and nobody
can use it on themselves — self-service goes through profile.php, which
demands the current password so a hijacked session cannot lock the owner
out.
Accounts provisioned during the SSO era have password_hash NULL. They get in
via the reset email, or by registering again with the same address:
auth_register_client() claims a passwordless row, but only a plain `client`
one, so knowing a staff address does not let anyone self-assign the role.
users.andmore_id is deliberately left in place. No code reads it any more, but
it records which central identity each row was linked to.
Deployment: apply database/migration_20260913_local_auth.sql (password_resets
+ auth_attempts) in the SAME sysadmin session that pulls this commit — the
forgot-password flow needs that table, and after the pull a login requires a
local password.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
