Add mutex protection and std::expected get() to all key stores (issue #36)

[dnovick]

What

Adds std::mutex protection and copy-under-lock get() functions returning std::expected<TView, CryptoError> to all seven key stores that were previously unprotected.

Why

Issue #36: concurrent import/destroy/get on the same key store had no synchronisation, creating data races. The PQC key stores already had a mutex but returned std::optional; this PR upgrades them to std::expected for consistency with the rest of the architecture.

Changes

• arm_asm/key_store.hpp, ia_asm/key_store.hpp: Added sym_store_mutex(), KeyView struct, key_store_get() → std::expected<KeyView, CryptoError> copying bytes under lock; import and destroy now locked
• arm_asm/ec_key_store.hpp: Added ec_store_mutex(), EcKeyView struct, ec_key_store_get() → std::expected<EcKeyView, CryptoError>
• arm_asm/rsa.hpp: Added rsa_store_mutex(), RsaKeyView struct, rsa_key_store_get() → std::expected<RsaKeyView, CryptoError>
• openssl/openssl_key_store.hpp: Added ossl_asym_store_mutex() and ossl_raw_store_mutex(); asym get() calls EVP_PKEY_up_ref() under lock (caller must EVP_PKEY_free()); raw get() copies bytes into OpenSslRawView; ossl_asym_store_alg() also locked
• arm_asm/pqc_key_store.hpp, psa_mbedtls/pqc_key_store.hpp: pqc_key_store_get_private/public upgraded from std::optional<PqcKeyView> to std::expected<PqcKeyView, CryptoError>
• arm_asm/arm_asm_backend.hpp, ia_asm/ia_asm_backend.hpp, arm_asm/hkdf.hpp, ia_asm/hkdf.hpp, openssl/openssl_backend.hpp: All call sites updated to use the new std::expected return types; OpenSSL asym call sites now call EVP_PKEY_free() at all exit paths
• safe-crypto-lib/crypto_error.hpp: Added CryptoErrorCode::InternalError for allocation/up_ref failures

Closes #36

[dnovick]

Blocking: this PR currently does not compile.

1. safe-crypto-lib-test/arm_asm_tests.hpp still calls the old pointer-out getter APIs after the PR changes the stores to expected-returning getters. The failing call sites include key_store_get(id, &key, &len) at lines 1264/1270/1277, ec_key_store_get(id, &curve, &kind, &key, &len) at lines 1330/1341/1349, and rsa_key_store_get(..., &kind, &bits, &key, &len) at lines 2308/2320/2329. These need to be updated to assert on the new std::expected return value instead of the old bool API.

2. The IA_ASM backend now calls rsa_key_store_get(id) as if the IA RSA store has the new view API, but providers/ia_asm/rsa.hpp still defines only the old pointer-out signature at line 78. CI fails in the IA_ASM compile job at the RSA export/sign/verify/encrypt/decrypt call sites in providers/ia_asm/ia_asm_backend.hpp (for example lines 310, 345, 551, 647, 783, and 805) with too-few-arguments errors. Either update the IA RSA key store to the new mutex-protected RsaKeyView/std::expected API or keep those backend call sites on the old signature.

Because every relevant CI job is failing at compile time, I did not run further local tests or merge.

[dnovick]

Blocking review on the updated #183: the previous API-mismatch blockers are fixed, but this still is not ready to merge.

1. CryptoError::message() now returns std::string_view, which breaks existing CLI string concatenation. A local scli build fails with invalid operands to binary expression (const char[...] and std::string_view) at the PQC CLI error paths, e.g. safe-crypto-cli/cmd_ml_kem.hpp:50, :90, :139, safe-crypto-cli/cmd_ml_dsa.hpp:49, :97, and safe-crypto-cli/cmd_slh_dsa.hpp:49, :100. Either keep message() returning const std::string&, or update these callers to explicitly build a std::string before concatenating.

2. The OpenSSL asymmetric store still has a consistency race between the key reference and the stored algorithm. ossl_asym_store_get() and ossl_asym_store_alg() each lock independently (providers/openssl/openssl_key_store.hpp:80-111), but sign/verify read them in separate calls (providers/openssl/openssl_backend.hpp:1191-1196 and :1251-1256). If another thread destroys/reimports the same slot between those calls, the operation can use an EVP_PKEY reference from the old slot together with the algorithm from the new slot. For issue Thread safety: symmetric, EC, RSA, and OpenSSL key stores have no synchronization #36, the OpenSSL asym getter should return a single view/reference object containing both the up-ref’d EVP_PKEY* and the alg captured under the same lock.

Verification: cmake -S /private/tmp/claude-crypto-pr183-worktree -B /private/tmp/claude-crypto-pr183-build -DSAFE_CRYPTO_ACTIVE_PROVIDER=PSA_MBEDTLS -DSAFE_CRYPTO_PQC=LIBOQS -DCMAKE_BUILD_TYPE=Debug configured successfully, but cmake --build /private/tmp/claude-crypto-pr183-build --target scli --parallel 4 fails on the std::string_view concatenation. GitHub currently shows OPENSSL / Debug failing as well, with some other checks still pending. Not merging.

[dnovick]

Not ready to merge yet. CI is now green and the earlier CryptoError::message() compile blocker is fixed, but the OpenSSL asymmetric key store still has the issue #36 consistency race I called out.

ossl_asym_store_get() and ossl_asym_store_alg() each take ossl_asym_store_mutex() separately (providers/openssl/openssl_key_store.hpp:80-110). Callers then read the key reference and algorithm in two separate operations, for example sign at providers/openssl/openssl_backend.hpp:1191-1196 and verify at providers/openssl/openssl_backend.hpp:1251-1256. If another thread destroys and reimports the same slot between those calls, the operation can hold an up-ref to the old EVP_PKEY while comparing against the new slot algorithm.

For the thread-safety issue this PR is closing, the OpenSSL asymmetric getter should capture both values under one lock, e.g. return a view/reference object containing the up-ref’d EVP_PKEY* plus the stored alg, and callers should use that single snapshot.

[dnovick]

Reviewed latest head 3220379. I do not see any remaining code blockers.

The previous issues look fixed:

• CryptoError::message() is back to returning std::string, and the scli target builds locally.
• OpenSSL asymmetric key operations now use ossl_asym_store_get_with_alg(), capturing the key pointer and stored algorithm under one lock before sign/verify/KEM checks.

Local verification passed:

• cmake --build /private/tmp/claude-crypto-pr183-build --target scli --parallel 4
• OpenSSL provider configure/build in /private/tmp/claude-crypto-pr183-openssl-build
• ctest --test-dir /private/tmp/claude-crypto-pr183-openssl-build --output-on-failure -E safe_crypto_cli_test_NOT_BUILT passed 529/529

GitHub CI is almost green: OpenSSL, PSA debug, PSA sanitize, TSan, IA compile, and all clang-tidy jobs have passed. ARM_ASM / Debug is still pending, so I would wait for that final required check before merging.