Discrepancies in abc/ — the foundational slice/buffer/net/poll/curl layer every other module builds on. Each entry is one ticket with its live state; it drops off once the ticket closes. The fix-and-land method is Issues. Memory-safety findings live under MEM; this topic is for behavioural/correctness gaps. The 2026-07-07 five-agent full-library review landed as ABC-005..018; every ticketed finding was verified against source first.
s_purge skips the swapped-in tail (LIVE via LSM/Y), $rm/$rm1 memmove the wrong count, u8sPop copies head yet sheds tail, sDrain breaks its own contract, $tailshift cannot compile, $feedf returns OK on end-of-template truncation.
bSplice mixes bytes and elements
(armed for any non-u8 T), bReMap shrink OOB (macOS path), bAcquire alignment + size-overflow math.
LSMMore unchecked memcpy
overflow + no-op sift-up, LSMSort silently loses data on roughly-sorted input, Yx $sort b8-comparator-through-qsort UB, PRO.h included by LSM.h.
nibble ("0g" decodes OK), OKscan matches "KO" not "OK" (0x518 vs 0x614), RON pad-feed breaks the Feed contract, ZINT-128 feeds unchecked.
huge records (bit-0x20 test) → stream desync; probe shift UB; leftover unaligned casts; missing u32 length guards.
JSON unescape emits CESU-8 its own validator rejects, decimal drains wrap u64 → OK with a wrong value, float parse truncates at 31 chars, i64Zig signed-overflow UB.
slice {h[0], *h}; verified empirically); RFC-legal empty reason-phrase rejected by the grammar.
NETInfo returns OK without copying
the result out, aNETAddress stack OOB on long hosts, bind walk aborts instead of trying the next address, CURL callbacks return OK on OOM (truncated body as HTTP 200), pol mask clobber, EINTR/EAGAIN collapse.
FILEFlushThreshold loses data on a
short write (Bate eats all), FILETrimMap truncates PAST off the file, FILE_BOOK grows over uninit memory vs its NULL sentinel, no O_CLOEXEC anywhere, FSW watches unremovable + kqueue fd leak.
wild DATA terminator (u64 wrap), PACKClose leaks on error, SLOG writer emits skip records its own reader cannot expand (64 vs 127), NESTFeed self-corrupts past its one-time room check.
bit-rot (ruling: revive or delete), HASHx converge compares a stale hash, MSET seeks with the wrong comparator, instantiation traps (X-clobber, leaked T/Key typedefs).
popc64 is the 32-bit builtin,
sha256empty unaligned u64 reads, MAIN exit-code truncation, SAN.h incomplete + undefs NULL, OK.h static rotating buffers.
so the net stack + fuzz targets never run in CI, LEX generator drifted from its committed output (fbreak vs goto) — blocks regens.
DNS name codec + NFA regex compiler are manual parsers, slice-from-arith cluster, naming/error-code drift, misc LOW list.
ABC-003 (UTF8_LEN[byte>>4] byte-length table shifted; multibyte lead bytes over-count) landed into the abc sub — 02a23dc3 (?/abc/.dogs) + pin cb71e65b. (BRO-004's incoherence was a separate bro bug, not this table — see ABC-003.)
ABC-004 (abc/BIT.h u1 bitmap + set algebra; BitUnset |=~ bug killed by construction; FILE_RW migrated; backs DOG-001's weavescope) landed into the abc sub — 7dafb8d7 + u64s rework 6e3b93f8 + pin 876315d1; the optional NFA class-bitmap migration was dropped as scope — see ABC-004.