Add ADR-018 (logging and log integrity)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docs/decisions/018-logging.md

@@ -0,0 +1,99 @@
+# ADR-018 — Logging and log integrity
+
+## Context
+
+boma wants all logs in one queryable store for troubleshooting, spotting issues over
+time, and detecting intrusions / malicious activity. ADR-002 commits in principle
+("logs shipped to a central location"; "active alerting wires AIDE/`auditd`/`fail2ban`/
+Suricata… ties to the Loki/Grafana effort"); CAPABILITIES lists Loki and `askari` (the
+off-site watchdog). Undecided: the architecture and the **integrity** question — an
+attacker who roots a host will try to clear logs to cover their tracks.
+
+The framing insight: the biggest anti-tampering win is that logs **leave the host in
+near-real-time** — once a line is in a store the attacker doesn't control, wiping the
+local copy is futile. How far to harden the central store is set by the threat model.
+
+## Decision
+
+1. **Threat model — opportunistic + blast-radius** (ADR-002 / accepted-risk R1). Not
+   forensic-grade.
+2. **All logs → an on-cluster Loki** — the single monitoring DB for troubleshooting +
+   trends. Near-real-time shipping already defeats per-host track-covering.
+3. **A security-relevant subset ALSO ships off-site to `askari`, write-only** —
+   tamper-resistant against full-cluster compromise, at bounded volume.
+4. **Skip WORM/object-lock** — accepted-risk R4; append-only push + off-site is the
+   proportionate control.
+5. **Disk-wear is a managed parameter** — media choice + bounded verbosity + tuned
+   retention + wearout monitoring.
+
+## Architecture
+
+- **Agent:** Grafana Alloy on every host, installed by the `base` role — reads journald
+  + container logs + security sources (`auditd`, `authpriv`, `fail2ban`, AIDE).
+- **Loki (cluster):** a `loki` service role on a docker_host; all logs; monolithic
+  single-binary mode; NVMe; bounded retention.
+- **Loki (`askari`):** the same role parameterised, in `offsite_hosts`; security subset
+  only, write-only, long retention, tiny volume.
+- **Grafana (cluster):** both Lokis as datasources (one pane queries both); dashboards
+  + the alerting ADR-002 calls for.
+
+## Data flow & the security subset
+
+Alloy writes everything to the cluster Loki and a filtered copy (a relabel/match stage
+tags security sources `security="true"`) to the `askari` Loki. Subset: `auditd`,
+`authpriv` (SSH/`sudo`), `fail2ban`, AIDE, **Suricata** (OPNsense isn't a `base` host —
+it syslog-forwards its alerts to the ingest point), and key container security events.
+
+**Write-only / append-only:** the `askari` push endpoint (`/loki/api/v1/push`) is
+mesh-only with a **push-only credential**; query/admin/delete APIs are not exposed to
+hosts. The push API has no edit/delete verb, so a compromised host can append but not
+read/edit/delete. The cluster Loki uses the same push-only credential. Alloy buffers
+(WAL) + retries across a brief outage.
+
+## Security, integrity & residual risks
+
+Defeats opportunistic track-covering (logs already off-host) and host-pivot-to-store
+(append-only, off-cluster). The security trail survives full-cluster compromise.
+Conscious residuals: append-only ≠ cryptographic WORM (root-on-`askari` could edit
+chunks — R4); a few-seconds un-shipped window; agent compromise can stop *future*
+shipping but not alter shipped history; **a host going silent is itself an alert**; a
+stolen push credential appends noise but can't delete; an `askari` outage buffers +
+flushes on reconnect.
+
+## Retention & disk-wear
+
+Estimates are intent-based until measured (like `/capacity-review`). Cluster Loki:
+bounded hot retention (~30–90 days). `askari` subset: long (~1 year+, ~5–25 GB/yr).
+Disk-wear rules: (1) log storage on NVMe/SSD or HDD, **never SD/USB flash**; (2) bounded
+verbosity at source (sane levels, selective access logging, a targeted `auditd`
+ruleset); (3) tuned Loki retention/compaction; (4) SSD **wearout/TBW** is a monitored
+metric (Proxmox wearout %, `node_exporter` smartmon) with an alert. Log storage is a
+tracked allocation in `docs/hardware/reference.md` (ADR-012).
+
+## Status
+
+Designed. **Authorable now:** this ADR + the ADR-002/CAPABILITIES/ADR-012/
+accepted-risks/STATUS/TODO reconciliations. **Deferred on the stack:** Alloy-in-`base`,
+the `loki`/`grafana` service roles, OPNsense syslog config, the push-only credential,
+and the live pipeline.
+
+## Dependencies
+
+`base` role + service-role machinery (unbuilt, STATUS.md); the running cluster +
+`askari` (`offsite_hosts`, ADR-016); OPNsense automation for Suricata syslog (ADR-007);
+the metrics stack (Prometheus / `node_exporter`) for SSD-wearout + log-silence alerting
+(sibling effort, TODO 3.6).
+
+## What was ruled out
+
+| Option | Reason |
+|---|---|
+| Everything off-site on `askari` (no on-cluster Loki) | The firehose is disk-hungry on a small VPS; keep volume where storage is cheap and send only the bounded security subset off-site. |
+| WORM / object-lock for all logs | Forensic-grade cost for an opportunistic threat model — YAGNI (R4). |
+| On-cluster-only (no off-site copy) | Doesn't survive compromise of the cluster Loki host; the security trail must be off-cluster + append-only. |
+| Volatile (RAM-only) journald to cut writes | Risks losing logs on crash before shipping; persistent-with-caps + real-time shipping is safer. |
+| Promtail / legacy agents | Alloy is the current unified Grafana collector and the V4-aligned choice (one agent for logs, later metrics). |
+
+See also: ADR-002 (security baseline — realised here), ADR-016 (mesh / `askari`),
+ADR-007 (OPNsense / `askari`), ADR-012 (hardware/capacity), ADR-004 (service-role
+standard), ADR-011 (health checks — distinct from this).