test(runner): a lease is proven to run its transport, under the full wall

The narrowing was reasoned, never measured. The evidence offered for it was
the self-test's egress row — a TCP connect issued from inside the statically
linked runner, which spawns nothing and so exercised the one path in a lease
that needs no executable at all. The claim and the measurement never touched,
which is why a review looking straight at it did not catch the fault.

Replace the reasoning with an A/B. A dynamically linked binary runs inside a
composed lease and must exit 0; the same command in the same argv with only
the system-core bind triples removed must exit non-zero. Neither arm proves
anything alone: a lone green stays green if the command would have run
anywhere, if the argv never applied, or if bwrap ignored the spliced tail.
The strip asserts it removed something, so a changed flag spelling fails
loudly instead of silently grading an unmodified lease.

Splicing the probe's tail drops `--sandboxed` with it, so those proofs cover
bwrap's MOUNTS only. The hardening half keeps the real tail intact and adds
`--transport=`, spawning from behind no_new_privs → landlock → seccomp, plus
an isolation test applying only the filesystem wall so a future refusal names
the layer that refused rather than reporting one composite failure.

The pin is deliberately `/usr/bin/env`, not `curl`: the kernel-lane image
ships no curl, and a proof that skips in the only environment continuous
integration runs it is not a proof. `curl` keeps its own test, gated on the
host carrying one.

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