[tracking] Unreproduced audit drafts (triage backlog, not findings)

Status: UNREPRODUCED — triage backlog, not findings

These are static observations from the same audit that produced the confirmed
issues. They were not attempted — the reproduction effort was timeboxed and
spent on the higher-severity candidates. Each is plausible and each has a
specific counter-hypothesis that could kill it. Please don't treat any of them
as a bug until someone has run it.

Filed as one issue so they are not lost, and so nobody spends the effort
rediscovering them. Reproducing any of these is cheap with the pattern the
confirmed ones used: drive Machine::system_call() directly on a fork and
compare the result against what Linux does.

Medium

• ppoll does not clamp its timeout on non-forked machines —
  system_calls.cpp:396. poll() clamps to 1 ms when !is_forked()
  (:330); ppoll has no equivalent. If that clamp exists to stop a master
  from blocking, ppoll is missing the same protection.

Low — wrong answer to the guest

• pwrite64 returns the requested length, not the pwritev64() result —
  system_calls.cpp:800. Wrong byte count on short writes.
• lseek truncates the 64-bit result to int for its error check —
  system_calls.cpp:286. A seek to an offset with bit 31 set in the low word
  reports failure, with a stale errno.
• accept4 returns without setting registers on policy denial —
  system_calls.cpp:1294. The guest sees a garbage success value.
• gettimeofday(NULL, …) writes the timeval to guest address 0 —
  system_calls.cpp:2151.
• flock reports LOCK_EX/LOCK_SH success without taking the lock —
  system_calls.cpp:1781.
• clock_nanosleep blocks the host thread on a guest-chosen timespec and
  reports failure as success — system_calls.cpp:2378.
• close() of a loaned (is_forked) fd returns success but keeps the
  entry — system_calls.cpp:177. Use-after-close then succeeds.

Low — arithmetic and bookkeeping

• mmap/mremap 64-bit addr + len wrap corrupts mmap-cache bookkeeping —
  system_calls.cpp:415. Medium confidence.
• m_total_fds_opened is uint16_t and wraps at 65536 — fds.cpp:138,
  defeating the embedder's total-fd limit.
• fcntl(F_DUPFD_CLOEXEC) hands an unchecked dup() to manage_duplicate() —
  system_calls.cpp:1890.
• translate() ignores KVM_TRANSLATE's valid bit — machine.cpp:370.
  Medium confidence; the only consequence found was a spurious
  MemoryException.

Low — no current caller

• copy_from_cstring() over-reads up to offset bytes past each page end —
  machine_utils.cpp:617. Zero callers today.
• relocate_section() / relocate_relr_section() wrapped relocation
  addresses, and elf_offset_array() count overflow — machine_elf.cpp:425.
  Embedder-API-only; dynamic_linking() has been a no-op since 864ad8d. This
  one was attempted in wave 2 and could not be reproduced.
• close() returns early without setting registers when free_fd_callback
  resets the VM — system_calls.cpp:170. Borderline out of scope, since it
  needs an embedder callback.
• ARM64 nested signal delivery grows SignalReturn::frames unboundedly —
  signals.cpp:87. Low confidence, and needs an AArch64 host.

Cosmetic, noted in passing

• rt_sigaction oldact on a never-set handler returns 0xFF…F0 instead of
  SIG_DFL — system_calls.cpp:608. Surfaced while reproducing the
  signal-handler part of #103.
• is_writable_path lacks the empty-path guard that is_readable_path
  has — fds.cpp:437 vs :420. Relevant to #96, where the policy layer is
  currently handed an empty path.
• thread_counter signed-int overflow → tid reuse — threads.cpp:159.
  Needs ~2^31 clones; noted on #105 since it is free to fix alongside the
  thread cap.

Start by choosing one unattempted candidate and read the named location in system_calls.cpp, fds.cpp, machine.cpp, machine_utils.cpp, machine_elf.cpp, signals.cpp, or threads.cpp. Drive Machine::system_call() directly on a fork and compare the result with Linux. Done means the behavior is reproduced and isolated, or the candidate is documented as not reproducible.