GHSA-5739-39v2-5754

### Impact

`RSACrypt::decryptWithRSA15()` (used by the `RSA1_5` key-encryption algorithm) implements RSAES-PKCS1-v1_5 decryption by inspecting the padding after RSADP and throwing `InvalidArgumentException` as soon as the padding is malformed. It does **not** implement the implicit-rejection countermeasure required by RFC 3447 §7.2.2 / RFC 8017 §7.1.2 (return a deterministic pseudo-random value of the expected length on padding failure and let the downstream step fail uniformly).

From a JWE caller this yields a Bleichenbacher/Marvin padding oracle: an attacker submitting adaptively crafted `encrypted_key` values can distinguish (a) padding rejected, (b) padding valid but wrong CEK length, and (c) padding valid and full AEAD executed — even though `JWEDecrypter` returns the same `false` in all cases — because each path performs a measurably different amount of work, amplifiable by enlarging the ciphertext (CWE-208 timing side channel). Enough adaptive queries can recover the wrapped CEK.

### Affected configurations

Applications that register `RSA1_5` in their decryption `AlgorithmManager` and hold an RSA private key.

### Patches

PKCS#1 v1.5 decryption now performs implicit rejection: on invalid padding (or unexpected recovered length) it returns a random CEK of the expected size selected in constant time, so the full content-decryption (AEAD) step always runs and fails uniformly, removing the observable difference between padding-valid and padding-invalid ciphertexts. Users are still strongly encouraged to migrate to `RSA-OAEP`.

### Workarounds

Prefer `RSA-OAEP`/`RSA-OAEP-256`; do not enable `RSA1_5` for untrusted tokens.

### References

- RFC 3447 §7.2.2 / RFC 8017 §7.1.2 - Bleichenbacher (1998); Marvin attack - CWE-208: Observable Timing Discrepancy

## Résolution

Un correctif a été préparé sur une branche dédiée basée sur `3.4.x`, avec des tests anti-régression dédiés (fork privé temporaire de cette advisory, PR #1).

**RSA1_5** — déchiffrement PKCS#1 v1.5 avec *implicit rejection* en temps constant (RFC 3447 §7.2.2) : un padding invalide n'est plus distinguable d'un padding valide, neutralisant l'oracle Bleichenbacher.

**Validation :** `php -l` OK, PHPUnit vert, aucune nouvelle erreur PHPStan introduite (différentiel nul vs `3.4.x`), aucun commentaire ajouté dans le code source. Après merge, cascade prévue `3.4.x → 4.0.x → 4.1.x`.