test(runner): pin "no executable in a lease" against the real lease argv

/write-integration-test's audit found §3's second defence unpinned. The
narrowed baseline drops /usr, /bin, /sbin, /lib and /lib64, so a lease
holds exactly one executable — the statically linked runner on its own
single-file bind. That is what makes a tool escaping the allowlist inert:
there is nothing to exec. Nothing in the suite failed if someone put one
of those trees back.

Asserted against the REAL argv from buildProbeArgv, with only the final
element swapped for a shell. A test that rebuilds the argv itself would
prove nothing here: argv-vs-list disagreement IS the failure class this
milestone kept hitting.

The audit's other finding needs no code. The resolver-symlink regression
that gave resolver=0 dns=0 egress=0 was ALREADY covered — "the resolver
check passes in an unmodified sandbox" drives the real argv and asserts
resolver_readable, and would have failed on it. It did not fire because
every test in this file opens with `if (builtin.os.tag != .linux) return
error.SkipZigTest`, so the decisive lane silently skips on macOS. The
coverage existed; the environment reading it as green did not.

This test compiles here but has NOT run: Linux + real bwrap only. The
test-integration-kernel lane is its proof.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>