fix: multiple OpenSSL 3.x memory leaks

[Koan-Bot]

Summary

Fixes three distinct memory leak families in OpenSSL >= 3.0.0 code paths:

• _get_key_parameters(): EVP_PKEY_get_bn_param() allocates new BIGNUMs (unlike pre-3.x getters which return internal const pointers), but cor_bn2sv() duplicates them via BN_dup() without freeing the originals — 8 BIGNUMs leaked per call
• verify(): XSRETURN_NO/XSRETURN_YES returned immediately, bypassing EVP_MD_free() and EVP_PKEY_CTX_free() cleanup — leaked on every verify call
• _new_key_from_parameters(): EVP_PKEY_CTX (pctx) was never freed on any path, test_ctx from EVP_PKEY_check() was never freed, and OSSL_PARAM_BLD/OSSL_PARAM were not freed in the public-key-only branch

Root cause

The OpenSSL 3.x migration changed ownership semantics: pre-3.x getter functions return internal pointers (caller must not free), while 3.x functions like EVP_PKEY_get_bn_param() allocate new objects (caller must free). The original migration missed several of these ownership changes.

Test plan

• All 278 tests pass (including bignum tests that exercise _get_key_parameters())
• Compiled with OpenSSL 3.6.1 on macOS
• CI will validate against OpenSSL 1.1.1 (bullseye), 3.0.x (bookworm), 3.4.x (trixie)

🤖 Generated with Claude Code

[timlegge]

looks like there is a newer version of this one in another PR

[Koan-Bot]

@timlegge — correct. PR #75 has been superseded by individual, more focused PRs:

• _new_key_from_parameters() cleanup → PR fix: free EVP_PKEY_CTX leak in _new_key_from_parameters() #87 (test_ctx leak) + PR fix: free OpenSSL 3.x resources in _new_key_from_parameters() #90 (pctx/params_build/params on all paths)
• verify() cleanup → PR fix: verify() resource leaks and dead cleanup code (OpenSSL 3.x) #81 (dead cleanup + error-path leaks)

However, one fix from this PR is not yet covered elsewhere: the _get_key_parameters() BIGNUM freeing. On OpenSSL 3.x, EVP_PKEY_get_bn_param() allocates new BIGNUMs (unlike pre-3.x RSA_get0_*() which returns internal pointers), so the 8 BIGNUMs (n, e, d, p, q, dmp1, dmq1, iqmp) need to be freed after use. I'll track this separately.

This PR can be closed once that remaining leak is addressed in its own PR.

[timlegge]

I did

s/>= 0x30000000L/>= 0x90000000L/g
s/< 0x30000000L/< 0x50000000L/g

to force the code to take the old version of the api and I have to change padding.t to do :

if $major lt '5.0' && $pad eq 'sslv23'

The old code path works fine.

Approvgving

[timlegge]

Should this be THROW(EVP_PKEY_fromdata_init(pctx) > 0);

[timlegge]

Good catch

[timlegge]

LGTM - The code duplication with the switch should be sorted later though