fix(enforce): stop advertising bpf_lsm once detached; pin guard for restart survival

[gnanirahulnutakki]

Summary

Epic A holistic review (#125), issues #121 and #124.

#121 (HIGH) — fail-open on guard-consumer death. If runGuardConsumer returned mid-run — a real error, or a clean ringbuf close caused by something other than the daemon's own shutdown watcher (e.g. an external force-detach) — its defer cleared d.policyMaps, but nothing ever moved activeTier off "bpf_lsm". Health kept reporting enforcement that no longer existed.

Fixed:

• The goroutine awaiting runGuardConsumer now calls degradeGuardTier on any non-shutdown exit (whether err is nil or not — both mean the guard is no longer attached).
• degradeGuardTier atomically moves activeTier to "none" and records the transition as a hash-chained tamper-audit entry — the same auditable evidence trail RunTamperAudit's periodic ticks already write to (_tamper/tamper_audit.jsonl), not just a log line that can be missed.
• activeTier reads/writes are now consistently mu-guarded (getActiveTier/setActiveTier) — this change adds a second concurrent writer to a field that was previously only ever set once at startup, so the existing unsynchronized access needed fixing to stay race-free under -race (which kernel-enforce.yml already runs).
• Deliberately does not attempt to retroactively stand up the seccomp fallback tier — that tier's supervisor/socket lifecycle is wired only at startup, and retrofitting a second start site would need its own careful lifecycle handling for a case this fix already reports correctly. "Loudly signal degradation," per the issue's own acceptance framing.

#124 (MED) — guard unpinned. A daemon restart detached the three LSM hooks and rebuilt every policy map empty (cgroup_managed, cgroup_op_policy, etc.), so a governed agent ran completely unenforced from process exit until (if ever) apply_policy was called again.

Fixed: LoadAndAttachProcessGuardEBPFPinned mirrors #99's tracepoint pinning — pins the three LSM links and the six policy-state maps (+ the enforce_events ringbuf) under /sys/fs/bpf/ardur/, all-or-nothing (never a partial reuse). A restart that finds every pin re-attaches to kernel state that was never actually detached: the LSM hooks kept enforcing the whole time, so there is nothing to "re-apply" — the pinned maps' contents are exactly what was last applied. If pinning is unavailable (bpffs not mounted, etc.), it falls back to a fresh load, the same accepted degradation #99 already has for the tracepoint.

Tests

• daemon_guard_degrade_test.go — proves health reflects the true tier after a simulated consumer death (pure state-transition logic, no real kernel needed): tier moves to none, the transition is hash-chained and appended to evidence, the no-op case (guard never actually reached bpf_lsm) doesn't spuriously fire, and a concurrent-access test exercises getActiveTier/setActiveTier/degradeGuardTier under -race.
• bpf_policy_apply_pinned_linux_test.go — pin-path logic (all 10 paths distinct, under the ardur bpffs namespace) and the "no pins exist yet" fallback (tryLoadPinnedGuardState failing cleanly at the ENOENT level, before ever touching BPF-LSM) — runs on real Linux in the bpf-generate CI job, no privileged VM needed.
• ardur-guard-smoke gained a third scenario, restart_survival_scenario.go (runs in kernel-smoke's virtme-ng VM): applies OP_EXEC:DENY through a pinned load, confirms EPERM, simulates a daemon restart (Close, then load again from the same pins), and asserts execve is still denied with no re-apply call in between. This can't pass by accident — a regression here (silent fallback to a fresh, unpinned attach) shows up as execve unexpectedly succeeding post-restart, not an ambiguous error.

Verification

• go build/vet/test -race ./... green on darwin
• go build/vet/test -race ./... green on a real Linux container (golang:1.26) — the new pinned-path tests actually executed there (confirmed via -v), not just cross-compiled
• ardur-guard-smoke binary builds and runs on real Linux, failing exactly where expected outside a privileged BPF-LSM kernel (confirms the control flow reaches the real loader correctly)
• Existing test suites (daemon, kernelcapture, guard-smoke build) all still green — no regressions

Refs: #121, #124, Epic A #63.