fix(leashd): degrade to proxy-only enforcement when cgroup path unavailable

[sixtoad]

Summary

On Docker Desktop with Kubernetes the agent container runs in a private cgroup namespace where /proc/self/cgroup reports 0::/. leash-entry's emitCgroupPath() skips / as invalid, so /leash/cgroup-path is never written, and leashd then FATALs at startup with cgroup path required (set --cgroup) — before any enforcement starts.

This is separate from #66 (which fixed the xt_cgroup network path): even with that fix, leashd can't start on Docker Desktop K8s because cgroup-path discovery fails first.

This change makes leash degrade gracefully instead of failing, following the same philosophy as #66 (cgroup unavailable → fall back + WARNING, never FATAL, preserve the boundary):

• cmd/leash-entry/main.go — when no scopable cgroup is found, log a WARNING and continue (don't abort the container) so leashd can start.
• internal/leashd/runtime.go — preFlight allows an empty cgroup path (proxy-only, warns); a non-empty path is still validated as before.
• internal/lsm/manager.go — UpdateRuntimeRules no-ops when there is no cgroup, so no cgroup-scoped BPF-LSM programs are attached (they'd have nothing valid to scope to).
• internal/assets/apply-{iptables,ip6tables,nftables}.sh — when there is no target cgroup, still isolate the control plane with a namespace-wide block (the same boundary as the existing network: fall back to blanket port block when cgroup isolation unavailable #66 cgroup fallback), so degraded mode stays safe.

Net: the kernel LSM layer (file/exec/connect) is disabled when the cgroup is undiscoverable, but the L7 MITM proxy (hostname/header/MCP) keeps enforcing and the control plane stays isolated — defense-in-depth minus the kernel layer, with a clear warning, rather than a hard failure.

Design note

This follows #66's pattern: automatic fallback + WARNING, no new flag. Since degraded mode is a reduction in enforcement, I'm happy to instead gate it behind an explicit opt-in flag (e.g. --allow-no-cgroup) if you'd prefer that posture.

Validation

• go build + go vet clean for internal/lsm, internal/leashd, cmd/leash-entry.
• New unit tests pass: LSM-manager no-op without a cgroup; preFlight allows a missing cgroup. All 14 existing preFlight tests still pass (non-empty invalid cgroup is still rejected).
• sh -n clean on all three netfilter scripts.
• Not yet validated end-to-end on Docker Desktop Kubernetes (no access to that environment). Would appreciate maintainer/CI verification there, plus a no-regression check on bare Linux. Happy to iterate.

Closes #67