Add ADR for embedded registry

docs/adrs/embedded-registry.md

+# Package spegel Distributed Registry Mirror
+
+Date: 2023-12-07
+
+## Status
+
+Accepted
+
+## Context
+
+Embedded registry mirror support has been on the roadmap for some time, to address multiple challenges:
+* Upstream registries may enforce pull limits or otherwise throttle access to images.
+* In edge scenarios, bandwidth is at a premium, if external access is available at all.
+* Distributing airgap image tarballs to nodes, and ensuring that images remain available, is an ongoing
+  hurdle to adoption.
+* Deploying an in-cluster registry, or hosting a registry outside the cluster, put significant
+  burden on administrators, and suffer from chicken-or-egg bootstrapping issues.
+
+An ideal embedded registry would have several characteristics:
+* Allow stateless configuration such that nodes can come and go at any time.
+* Integrate into existing containerd registry mirror support.
+* Integrate into existing containerd image stores such that an additional copy of layer data is not required.
+* Use existing cluster authentication mechanisms to prevent unauthorized access to the registry.
+* Operate with minimal added CPU and memory overhead.
+
+## Decision
+
+* We will embed spegel within K3s, and use it to host a distributed registry mirror.
+* The distributed registry mirror will be enabled cluster-wide via server CLI flag.
+* Selection of upstream registries to mirror will be implemented via the existing `registries.yaml`
+configuration file.
+* The registry API will be served via HTTPS on every node's private IP at port 6443. On servers this will
+use the existing supervisor listener; on agents a new listener will be created for this purpose.
+* The default IPFS/libp2p port of 5001 will be used for P2P layer discovery.
+* Access to the registry API and P2P network will require proof of cluster membership, enforced via
+client certificate or preshared key.
+* Hybrid/multicloud support is out of scope; when the distributed registry mirror is enabled, cluster
+members are assumed to be directly accessible to each other via their internal IP on the listed ports.
+
+## Consequences
+
+* The size of our self-extracting binary and Docker images increase by several megabytes.
+* We take on the support burden of keeping spegel up to date, and supporting its use within K3s.