Problem

The final descriptor-sealing pass first tries
close_range(..., CLOSE_RANGE_CLOEXEC), then enumerates /proc/self/fd,
then falls back to scanning descriptor numbers 0 through 1023.

When close_range() is unavailable or denied and mount/chroot setup hides
/proc, inherited descriptors above 1023 are not marked close-on-exec and can
survive into the jailed target.

Change

Open /proc/self/fd before mount namespace setup and retain that directory
stream until the final descriptor-sealing pass. The pass can then enumerate the
complete final descriptor set even when the configured jail does not expose
/proc.

The existing order is preserved:

1. use close_range() when available;
2. enumerate the retained or current proc directory;
3. retain the bounded fallback if neither mechanism is available.

The retained stream is closed by its owner on every containProc() return path.

Verification

• Baseline with close_range() forced to return EPERM, --disable_proc, and
  an inherited descriptor 1500: the target can observe descriptor 1500.
• Patched build under the same conditions: descriptor 1500 is closed before
  target execution.
• Explicit --pass_fd 1500 remains preserved.
• A closed --pass_fd number reused by the retained proc directory is released
  by the stream owner before target execution.
• Standalone-once, standalone-execve, PID namespace, --execute_fd, and
  listen-mode paths were exercised.
• The repository Docker build completes with -Werror.

This is a hardening change for the process-isolation boundary and does not
change the behavior of explicitly passed descriptors.