[Task]: Implement the core create→start→kill→delete lifecycle test

[michael-duren]

What needs to be done

Implement the first end-to-end acceptance test that drives the real box binary
through the full lifecycle against a real bundle, asserting the behavior our unit
tests deliberately can't reach.

Using the harness (#18) and bundle fixture
(#19), the test should:

1. Build (or locate) ./bin/box and a fresh bundle running a long-lived process
   (e.g. sleep).
2. box create <id> --bundle <dir> → assert box state <id> reports created
   (this exercises the Init ↔ Reexec unix-socket handshake).
3. box start <id> → assert status running and that the user process is
   actually alive (covers the start handshake + syscall.Exec).
4. box kill <id> (e.g. SIGTERM/SIGKILL) → assert status stopped and the
   process is gone (covers real signal delivery).
5. box delete <id> → assert the state and runtime directories are removed and
   box state <id> now errors (covers process kill + cleanup).

Also add the obvious negative cases that are only meaningful end-to-end, e.g.
start on an unknown id, delete of a running container without --force.

This is where the methods marked TODO(tdd) / skipped in
internal/container/container_test.go (Init, Reexec, and the happy paths of
Start/Kill/Delete) get their coverage.

Acceptance criteria

• An acceptance test drives create → start → state → kill → delete against the real binary
• Status transitions asserted at each step (created → running → stopped)
• The user process is verified running after start and gone after kill
• delete is verified to remove state + runtime files
• At least two negative/end-to-end-only cases included
• The test skips cleanly when prerequisites (privileges/rootfs/Linux) are missing

Parent / tracking issue

Part of #17.

Blocked by

• [Task]: Research & design the acceptance test harness #18
• [Task]: Build a reusable test bundle + rootfs fixture #19

Rough size

L — a few days