High priority findings

• High: buffered stdout/stderr writes report overflow but still write past the configured end. That happens in src/zen_io.c:88, src/zen_io.c:143, and src/
  zen_io.c:177. Those branches should return immediately after the error instead of continuing into memcpy.
• High: the binding executor has two fixed-buffer bugs. line_alloc() allocates B64decoded_len(len) bytes and then writes a terminator at line[reallen] in src/
  zencode-exec.c:61, so it needs +1. The config path also appends ",logfmt=json" with strcat() in src/zencode-exec.c:123 after a truncation check in src/zencode-
  exec.c:114 that can never catch a too-long fgets() result.
• High: the X.509 helpers treat DER slices as C strings. src/zen_x509.c:138 and src/zen_x509.c:279 use %s on binary certificate data, and src/zen_x509.c:283 uses
  sprintf() unnecessarily. This is an over-read/overflow risk and also misparses embedded NULs and ASN.1 long-form lengths.
• High: native runtime state is still effectively global. src/zenroom.c:62, src/zenroom.c:65, src/zen_octet.h:69, and src/zen_error.h:32 hide allocator/context
  ownership behind global singletons and macro-redefined malloc/free. That hurts readability, makes the VM non-reentrant, and complicates maintenance.
• High: the logging helper macro is unsafe. src/zen_error.c:60 passes a va_list to _err() as if it were normal variadic arguments, and callers like src/
  zen_error.c:185 return early without va_end(). NULL-context logging is therefore undefined behavior today.
• High: both Zcash import guards are inverted. src/zen_ecp.c:417 and src/zen_ecp2.c:676 use SAFE_GOTO with a predicate that describes the invalid case, so valid
  headers fall into the error path.
• Medium-high: zen_init() has several error-path ownership problems. src/zenroom.c:176 does not check malloc, src/zenroom.c:217 leaks ZZ on config-parse failure,
  src/zenroom.c:232 does not check RNG allocation, and src/zenroom.c:288 can return after Lua state creation without lua_close().
• Medium: RNG-seed validation skips the last 4 hex characters. In src/zen_config.c:124 len is reduced by 4, then the validation loop starts at 4 and stops at
  that reduced bound in src/zen_config.c:131. The copy at src/zen_config.c:140 still takes the full payload.
• Medium: time_arg() ignores its n parameter for scalar inputs and always reads slot 1 in src/zen_time.c:89 and src/zen_time.c:110. That is a correctness bug and
  makes the helper misleading.
• Medium: big:bytes() is wrong for values whose bit length is not divisible by 8. src/zen_big.c:870 computes ceil(_bitsize(d)/8) after integer division, so 9
  bits still reports 1 byte.
• Medium: class metatable naming can overflow. src/lua_functions.c:86 seeds a 512-byte buffer with "zenroom.", then src/lua_functions.c:87 appends up to 511 more
  bytes with strncat().
• Medium: the Longfellow bridge silently truncates inputs and passes octet buffers where the external API expects C strings. See src/zen_longfellow.c:140, src/
  zen_longfellow.c:147, and the direct const char * calls at src/zen_longfellow.c:271 and src/zen_longfellow.c:345. This boundary needs explicit length checks
  and NUL-terminated staging buffers.
• Medium: api_hash has an unsafe cleanup path. tmp.val is never initialized before the first possible goto end in src/api_hash.c:159, but it is conditionally
  freed at src/api_hash.c:210.
• Medium: zen_parse.c still relies on static scratch storage in src/zen_parse.c:39. That is easy to replace with a stack buffer and would remove another hidden
  shared state point.
• Medium: zen_io.c has a dead Emscripten branch due to a typo in src/zen_io.c:182 (_EMSCRIPTEN instead of EMSCRIPTEN).

Lower-Priority Cleanup

• Low: src/zen_hash.h:35 and src/zen_hash.h:38 define _SHA3_256 twice.
• Low: src/zen_hash.c:75 and the repeated strncpy() calls below it never force NUL termination of h->name, and the local ht is effectively dead.
• Low: src/randombytes.c:218 and src/randombytes.c:227 leak the Linux fd on some failure exits. src/randombytes.c:109 still uses legacy CryptoAPI on Windows.
• Low: src/encoding.c:589 and src/encoding.c:610 implement BIP39 lookup with fixed magic sizes and a linear 2048-word scan. It is correct enough, but not
  especially clear or efficient.
• Low: there is repeated near-identical code for raw writes/logging in src/zen_io.c:51 and src/zen_error.c:46. One shared helper would simplify maintenance.