Hard-crashed egress leaves stale iptables/nft rules and a zombie mitmdump
holding port 18081; restarting the container then accumulates duplicate
rules and the new mitmdump fails to bind, sending the in-process mitm
watchdog (PR opensandbox-group#942) into a retry loop. This change keeps the egress
process under a dedicated supervisor so restarts are deterministic and
the dirty state is reset on every launch and exit.

components/internal/supervisor: new shared single-worker supervisor.
Exponential backoff with jitter, pre-start / post-exit hooks (failures
non-fatal), crashloop circuit breaker, JSONL event log. SIGTERM is
forwarded to the worker with a configurable grace period before SIGKILL.
Includes unit + integration tests using a re-exec'd test binary as a
fake child.

components/internal/cmd/supervisor: opensandbox-supervisor binary built
from the same module; flag-driven, no new external deps.

components/egress/scripts/cleanup.sh: best-effort, idempotent reset of
the iptables DNS REDIRECT rules, transparent-HTTP rules, the
nftables `opensandbox` table, and stray mitmdump processes. Hard
contract: never exit non-zero so a misbehaving cleanup cannot block
restarts.

components/egress/Dockerfile: builds and installs the supervisor and
the cleanup script alongside the egress binary under
/opt/opensandbox-egress/, then switches the ENTRYPOINT to run the
supervisor with cleanup as both pre-start and post-exit, grace 20s to
cover the egress-internal shutdown budget.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

PR review (opensandbox-group#951, Codex P1) caught a real security
regression: the post-exit cleanup hook was tearing down the iptables DNS
REDIRECT rules and the `inet opensandbox` nft table before the supervisor
slept for backoff and relaunched egress. Because the egress sidecar shares
its network namespace with the workload it is meant to filter, that window
left the workload with unfiltered egress instead of the stale default-deny
rules continuing to protect it. With a worst-case crashloop budget of
10 launches over 5 minutes, that window can stretch to minutes.

The fix is to leave netfilter state alone between runs:

- Drop the post-exit hook entirely. The backoff window now keeps the
  previous run's enforcement rules in place.
- Slim cleanup.sh to mitmdump-reaping only. iptables rule accumulation
  across many restarts is a slower-burn drift that egress's own
  SetupRedirect tolerates (first match wins); the nftables manager
  already prepends `delete table` to its ruleset script, so ApplyStatic
  is idempotent. Neither needs hook intervention.
- Keep the pre-start mitmdump kill so the new egress can bind the
  transparent-MITM listen port without colliding with an orphan.

(Codex P2 — zombie reaping when supervisor is PID 1 — is intentionally
not addressed in this commit; it does not gate the security fix.)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>