Update renovate/renovate Docker tag to v43.233.4

Reviewed-on: #331

.woodpecker/coredns-build.yaml

@@ -0,0 +1,46 @@
+when:
+  - event: push
+    path:
+      include:
+        - mikrotik/coredns/**
+
+steps:
+  - name: Get registry creds from OpenBao
+    image: quay.io/openbao/openbao:2.5.5
+    environment:
+      VAULT_ADDR: https://openbao.lumpiasty.xyz:8200
+      ROLE_ID:
+        from_secret: renovate_role_id
+      SECRET_ID:
+        from_secret: renovate_secret_id
+    commands:
+      - bao write -field token auth/approle/login
+          role_id=$ROLE_ID
+          secret_id=$SECRET_ID > /woodpecker/.vault_id
+      - export VAULT_TOKEN=$(cat /woodpecker/.vault_id)
+      - 'printf "PLUGIN_USERNAME=%s\n" "$(bao kv get -mount secret -field REGISTRY_USERNAME container-registry)" > /woodpecker/registry.env'
+      - 'printf "PLUGIN_PASSWORD=%s\n" "$(bao kv get -mount secret -field REGISTRY_PASSWORD container-registry)" >> /woodpecker/registry.env'
+
+  - name: Build and push
+    image: woodpeckerci/plugin-docker-buildx:6.1.0
+    privileged: true
+    settings:
+      registry: gitea.lumpiasty.xyz
+      repo: gitea.lumpiasty.xyz/lumpiasty/coredns-mikrotik
+      platforms: linux/arm64
+      tags:
+        - latest
+        - ${CI_COMMIT_SHA:0:8}
+      dockerfile: mikrotik/coredns/Dockerfile
+      context: mikrotik/coredns/
+      env_file: /woodpecker/registry.env
+
+  - name: Invalidate OpenBao token
+    image: quay.io/openbao/openbao:2.5.5
+    environment:
+      VAULT_ADDR: https://openbao.lumpiasty.xyz:8200
+    commands:
+      - export VAULT_TOKEN=$(cat /woodpecker/.vault_id)
+      - bao write -f auth/token/revoke-self
+    when:
+      - status: [success, failure]

.woodpecker/flux-reconcile-source.yaml

@@ -6,7 +6,7 @@ skip_clone: true
 
 steps:
   - name: Get kubernetes access from OpenBao
-    image: quay.io/openbao/openbao:2.5.4
+    image: quay.io/openbao/openbao:2.5.5
     environment:
       VAULT_ADDR: https://openbao.lumpiasty.xyz:8200
       ROLE_ID:
@@ -20,7 +20,7 @@ steps:
       - export VAULT_TOKEN=$(cat /woodpecker/.vault_id)
       - bao write -format json -f /kubernetes/creds/flux-reconcile > /woodpecker/kube_credentials
   - name: Construct Kubeconfig
-    image: alpine/k8s:1.36.1
+    image: alpine/k8s:1.36.2
     environment:
       KUBECONFIG: /woodpecker/kubeconfig
     commands:
@@ -41,7 +41,7 @@ steps:
     commands:
       - flux reconcile source git flux-system
   - name: Invalidate OpenBao token
-    image: quay.io/openbao/openbao:2.5.4
+    image: quay.io/openbao/openbao:2.5.5
     environment:
       VAULT_ADDR: https://openbao.lumpiasty.xyz:8200
     commands:

.woodpecker/renovate.yaml

@@ -6,7 +6,7 @@ skip_clone: true
 
 steps:
   - name: Get renovate token from OpenBao
-    image: quay.io/openbao/openbao:2.5.4
+    image: quay.io/openbao/openbao:2.5.5
     environment:
       VAULT_ADDR: https://openbao.lumpiasty.xyz:8200
       ROLE_ID:
@@ -21,7 +21,7 @@ steps:
       - bao kv get -mount secret -field RENOVATE_TOKEN renovate > /woodpecker/renovate_token
       - bao kv get -mount secret -field GITHUB_COM_TOKEN renovate > /woodpecker/github_com_token
   - name: Run Renovate
-    image: renovate/renovate:43.210.2
+    image: renovate/renovate:43.233.4
     environment:
       RENOVATE_AUTODISCOVER: "true"
       RENOVATE_ENDPOINT: https://gitea.lumpiasty.xyz/api/v1
@@ -34,7 +34,7 @@ steps:
       - export GITHUB_COM_TOKEN=$(cat /woodpecker/github_com_token)
       - /usr/local/sbin/renovate-entrypoint.sh renovate
   - name: Invalidate OpenBao token
-    image: quay.io/openbao/openbao:2.5.4
+    image: quay.io/openbao/openbao:2.5.5
     environment:
       VAULT_ADDR: https://openbao.lumpiasty.xyz:8200
     commands:

ansible/roles/routeros/tasks/addressing.yml

@@ -40,16 +40,18 @@
       - address: 2001:470:70:dd::2/64
         advertise: false
         interface: sit1
-      - address: ::ffff:ffff:ffff:ffff/64
-        from-pool: pool1
+      # Static instead of from-pool: pool allocation is dynamic (first free /64,
+      # e.g. ...:0::/64) which made the RDNSS address advertised in ND config
+      # point at a nonexistent router address. HE prefix is static, so static
+      # per-VLAN addressing is deterministic and matches docs/network.md.
+      - address: 2001:470:61a3:9:ffff:ffff:ffff:ffff/64
         interface: vlan2
       - address: 2001:470:61a3:500:ffff:ffff:ffff:ffff/64
         interface: containers
       - address: 2001:470:61a3:100::1/64
         advertise: false
         interface: vlan4
-      - address: ::ffff:ffff:ffff:ffff/64
-        from-pool: pool1
+      - address: 2001:470:61a3:a:ffff:ffff:ffff:ffff/64
         interface: vlan5
       - address: 2001:470:61a3:600::1/64
         advertise: false

ansible/roles/routeros/tasks/base.yml

@@ -65,6 +65,9 @@
       - bridge: containers
         interface: veth-tailscale
         comment: Tailscale container interface
+      - bridge: containers
+        interface: veth-coredns
+        comment: CoreDNS container interface
       - bridge: bridge1
         interface: ether1
         pvid: 2
@@ -152,24 +155,9 @@
     handle_absent_entries: remove
     handle_entries_content: remove_as_much_as_possible
 
-- name: Configure DHCP networks
-  community.routeros.api_modify:
-    path: ip dhcp-server network
-    data:
-      - address: 192.168.0.0/24
-        dns-server: 192.168.0.1
-        gateway: 192.168.0.1
-      - address: 192.168.255.0/24
-        dns-none: true
-        gateway: 192.168.255.10
-      - address: 192.168.5.0/24
-        dns-server: 192.168.5.1
-        gateway: 192.168.5.1
-    handle_absent_entries: remove
-    handle_entries_content: remove_as_much_as_possible
-
-# TODO: IPv6 pools are useful when we have dynamic prefix, but we don't
-# We can remove it now
+# Pool is no longer referenced — vlan2/vlan5 now use static addresses
+# (addressing.yml) so the RDNSS addresses in ND config are deterministic.
+# Kept defined for one run after migration; safe to delete afterwards.
 - name: Configure IPv6 pools
   community.routeros.api_modify:
     path: ipv6 pool
@@ -188,7 +176,9 @@
     values:
       allow-remote-requests: true
       cache-size: 20480
-      servers: 1.1.1.1,1.0.0.1,2606:4700:4700::1111,2606:4700:4700::1001
+      # CoreDNS container: plain forwarder with selective AAAA suppression.
+      # Forwards upstream to 1.1.1.1/8.8.8.8.
+      servers: 172.20.0.3
 
 - name: Configure DNS static entries
   community.routeros.api_modify:
@@ -199,6 +189,12 @@
         forward-to: 100.100.100.100
         match-subdomain: true
         comment: Tailscale MagicDNS
+      # Do NOT add a lumpiasty.xyz FWD entry here. RouterOS FWD entries return
+      # NOERROR with an empty answer instead of relaying NXDOMAIN, which breaks
+      # getaddrinfo search-domain processing (ENOTFOUND for valid names in k8s
+      # pods). Our own zone is handled in the CoreDNS Corefile (lumpiasty.xyz
+      # server block, AAAA kept) which relays rcodes correctly.
+      # See docs/coredns.md.
     handle_absent_entries: remove
     handle_entries_content: remove_as_much_as_possible
 
@@ -244,6 +240,22 @@
     handle_absent_entries: remove
     handle_entries_content: remove_as_much_as_possible
 
+- name: Configure DHCP networks
+  community.routeros.api_modify:
+    path: ip dhcp-server network
+    data:
+      - address: 192.168.0.0/24
+        dns-server: 192.168.0.1
+        gateway: 192.168.0.1
+      - address: 192.168.255.0/24
+        dns-none: true
+        gateway: 192.168.255.10
+      - address: 192.168.5.0/24
+        dns-server: 192.168.5.1
+        gateway: 192.168.5.1
+    handle_absent_entries: remove
+    handle_entries_content: remove_as_much_as_possible
+
 - name: Configure IPv6 ND defaults
   community.routeros.api_find_and_modify:
     ignore_dynamic: false
@@ -252,3 +264,21 @@
       default: true
     values:
       advertise-dns: true
+
+# RDNSS (RFC 8106): advertise an IPv6 DNS server in RAs so dual-stack clients
+# have an IPv6 resolver. Points at the router's per-VLAN IPv6 address; RouterOS
+# DNS forwards to CoreDNS. No pref64 — NAT64 has been removed (see docs/coredns.md);
+# AAAA suppression now happens in CoreDNS, no client-side translation needed.
+- name: Configure IPv6 ND per-interface (RDNSS)
+  community.routeros.api_modify:
+    path: ipv6 nd
+    data:
+      # advertise-dns must be explicitly enabled — RouterOS creates new ND
+      # entries with advertise-dns=no, which suppresses the RDNSS option
+      # entirely even when a static dns= list is configured.
+      - interface: vlan2
+        advertise-dns: true
+        dns: 2001:470:61a3:9:ffff:ffff:ffff:ffff
+      - interface: vlan5
+        advertise-dns: true
+        dns: 2001:470:61a3:a:ffff:ffff:ffff:ffff

ansible/roles/routeros/tasks/containers.yml

@@ -20,15 +20,15 @@
     data:
       - dst: /var/lib/tailscale
         list: tailscale_state
-        src: tailscale/state
+        src: /tailscale/state
     handle_absent_entries: remove
     handle_entries_content: remove_as_much_as_possible
 
-- name: Configure tailscale container
+- name: Configure containers
   community.routeros.api_modify:
     path: container
     data:
-      - dns: 172.17.0.1
+      - dns: 172.20.0.1
         interface: veth-tailscale
         logging: true
         mountlists: tailscale_state
@@ -36,5 +36,12 @@
         remote-image: gitea.lumpiasty.xyz/lumpiasty/mikrotik-tailscale:stable
         root-dir: tailscale/root
         start-on-boot: true
+      - dns: 172.20.0.1
+        interface: veth-coredns
+        logging: true
+        name: coredns
+        remote-image: gitea.lumpiasty.xyz/lumpiasty/coredns-mikrotik:latest
+        root-dir: coredns/root
+        start-on-boot: true
     handle_absent_entries: remove
     handle_entries_content: remove_as_much_as_possible

ansible/roles/routeros/tasks/firewall.yml

@@ -1,8 +1,56 @@
 ---
+- name: Configure WAN connection marking
+  community.routeros.api_modify:
+    path: ip firewall mangle
+    data:
+      - action: mark-connection
+        chain: forward
+        connection-state: new
+        new-connection-mark: wan-gpon
+        out-interface: pppoe-gpon
+        passthrough: true
+        comment: Mark connections going out GPON
+      - action: mark-connection
+        chain: forward
+        connection-state: new
+        new-connection-mark: wan-lte
+        out-interface: vlan6
+        passthrough: true
+        comment: Mark connections going out LTE
+    handle_absent_entries: remove
+    handle_entries_content: remove_as_much_as_possible
+    ensure_order: true
+
 - name: Configure IPv4 firewall filter rules
   community.routeros.api_modify:
     path: ip firewall filter
     data:
+      - action: reject
+        chain: forward
+        connection-mark: wan-gpon
+        out-interface: vlan6
+        protocol: tcp
+        reject-with: tcp-reset
+        comment: Fast-fail TCP connections that shifted from GPON to LTE
+      - action: reject
+        chain: forward
+        connection-mark: wan-gpon
+        out-interface: vlan6
+        reject-with: icmp-network-unreachable
+        comment: Fast-fail non-TCP connections that shifted from GPON to LTE
+      - action: reject
+        chain: forward
+        connection-mark: wan-lte
+        out-interface: pppoe-gpon
+        protocol: tcp
+        reject-with: tcp-reset
+        comment: Fast-fail TCP connections that shifted from LTE to GPON
+      - action: reject
+        chain: forward
+        connection-mark: wan-lte
+        out-interface: pppoe-gpon
+        reject-with: icmp-network-unreachable
+        comment: Fast-fail non-TCP connections that shifted from LTE to GPON
       - action: fasttrack-connection
         chain: forward
         connection-state: established,related

ansible/roles/routeros/tasks/routing.yml

@@ -12,15 +12,44 @@
         scope: 30
         suppress-hw-offload: false
         target-scope: 10
-      - disabled: false
+      - comment: GPON Monitor 1
+        disabled: false
+        distance: 1
+        dst-address: 1.0.0.1/32
+        gateway: pppoe-gpon
+        routing-table: main
+        scope: 10
+        suppress-hw-offload: false
+        target-scope: 10
+      - comment: GPON Monitor 2
+        disabled: false
+        distance: 1
+        dst-address: 8.8.4.4/32
+        gateway: pppoe-gpon
+        routing-table: main
+        scope: 10
+        suppress-hw-offload: false
+        target-scope: 10
+      - comment: GPON Default 1
+        disabled: false
         distance: 1
         dst-address: 0.0.0.0/0
-        gateway: pppoe-gpon
+        gateway: 1.0.0.1
+        check-gateway: ping
         routing-table: main
         scope: 30
         suppress-hw-offload: false
-        target-scope: 10
-        vrf-interface: pppoe-gpon
+        target-scope: 11
+      - comment: GPON Default 2
+        disabled: false
+        distance: 2
+        dst-address: 0.0.0.0/0
+        gateway: 8.8.4.4
+        check-gateway: ping
+        routing-table: main
+        scope: 30
+        suppress-hw-offload: false
+        target-scope: 11
     handle_absent_entries: remove
     handle_entries_content: remove_as_much_as_possible
 
@@ -32,6 +61,7 @@
         distance: 1
         dst-address: 2000::/3
         gateway: 2001:470:70:dd::1
+        check-gateway: ping
         scope: 30
         target-scope: 10
       - comment: Tailnet

ansible/roles/routeros/tasks/wan.yml

@@ -8,7 +8,9 @@
         keepalive-timeout: 2
         name: pppoe-gpon
         password: "{{ routeros_pppoe_password }}"
-        use-peer-dns: true
+        # Using CoreDNS container with DNS64
+        use-peer-dns: false
+        add-default-route: false
         user: "{{ routeros_pppoe_username }}"
     handle_absent_entries: remove
     handle_entries_content: remove_as_much_as_possible
@@ -37,5 +39,10 @@
         mac-address: 7E:7E:A1:B1:2A:7B
         name: veth-tailscale
         comment: Tailscale container
+      - address: 172.20.0.3/24
+        dhcp: false
+        gateway: 172.20.0.1
+        name: veth-coredns
+        comment: CoreDNS container
     handle_absent_entries: remove
     handle_entries_content: remove_as_much_as_possible

apps/authentik/release.yaml

@@ -18,7 +18,7 @@ spec:
   chart:
     spec:
       chart: authentik
-      version: 2026.5.2
+      version: 2026.5.3
       sourceRef:
         kind: HelmRepository
         name: authentik

apps/gitea/gitea-shared-volume.yaml

@@ -7,7 +7,7 @@ metadata:
   name: gitea-shared-storage-lvmhdd
   namespace: openebs
 spec:
-  capacity: 10Gi
+  capacity: "21474836480"
   ownerNodeID: anapistula-delrosalae
   shared: "yes"
   thinProvision: "no"
@@ -20,7 +20,7 @@ metadata:
   name: gitea-shared-storage-lvmhdd
 spec:
   capacity:
-    storage: 10Gi
+    storage: 20Gi
   accessModes:
     - ReadWriteOnce
   persistentVolumeReclaimPolicy: Retain
@@ -41,6 +41,6 @@ spec:
     - ReadWriteOnce
   resources:
     requests:
-      storage: 10Gi
+      storage: 20Gi
   storageClassName: hdd-lvmpv
   volumeName: gitea-shared-storage-lvmhdd

apps/llama/configs/config.yaml

@@ -14,12 +14,13 @@ macros:
   qwen35_35b_heretic_mmproj: "--mmproj-url https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/resolve/main/mmproj-F16.gguf --mmproj /root/.cache/llama.cpp/unsloth_Qwen3.5-35B-A3B-GGUF_mmproj-F16.gguf"
   qwen35_4b_heretic_mmproj: "--mmproj-url https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/mmproj-F16.gguf --mmproj /root/.cache/llama.cpp/unsloth_Qwen3.5-4B-GGUF_mmproj-F16.gguf"
   gemma4_sampling: "--temp 1.0 --top-p 0.95 --top-k 64 -ctk q4_0 -ctv q4_0"
+  gemma4_nothink_sampling: "--temp 1.0 --top-p 0.95 --top-k 64 -ctk q4_0 -ctv q4_0 --reasoning off"
 
 hooks:
   on_startup:
     preload:
       - "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"
-      - "whisper-small"
+      - "parakeet-tdt_ctc-1.1b"
 
 # matrix replaces groups (they are mutually exclusive).
 # The small 0.8B model runs alongside any LLM.
@@ -27,7 +28,7 @@ hooks:
 matrix:
   vars:
     q8: "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"
-    stt: "whisper-small"
+    stt: "parakeet-tdt_ctc-1.1b"
     flux: "flux2-klein-4b:Q4_K_M"
     coder: "Qwen3-Coder-Next-GGUF:Q4_K_M"
     q35t: "Qwen3.5-35B-A3B-GGUF:Q4_K_M"
@@ -38,10 +39,24 @@ matrix:
     q4nt: "Qwen3.5-4B-GGUF-nothink:Q4_K_M"
     q4ht: "Qwen3.5-4B-heretic-GGUF:Q4_K_M"
     q4hnt: "Qwen3.5-4B-heretic-GGUF-nothink:Q4_K_M"
-    g26xl: "gemma-4-26B-A4B-it:UD-Q4_K_XL"
-    g26q2: "gemma-4-26B-A4B-it:UD-Q2_K_XL"
-    ge4xl: "unsloth/gemma-4-E4B-it-GGUF:UD-Q4_K_XL"
-    ge2xl: "unsloth/gemma-4-E2B-it-GGUF:UD-Q4_K_XL"
+    g26xl: "gemma-4-26B-A4B-it-qat:UD-Q4_K_XL"
+    g26xlnt: "gemma-4-26B-A4B-it-qat-nothink:UD-Q4_K_XL"
+    g26mtp: "gemma-4-26B-A4B-it-qat-mtp:UD-Q4_K_XL"
+    g26mtpnt: "gemma-4-26B-A4B-it-qat-mtp-nothink:UD-Q4_K_XL"
+    g26ht: "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF:UD-Q4_K_XL"
+    g26hnt: "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF-nothink:UD-Q4_K_XL"
+    g26hmtp: "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF-mtp:UD-Q4_K_XL"
+    g26hmnt: "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF-mtp-nothink:UD-Q4_K_XL"
+    ge4qat: "unsloth/gemma-4-E4B-it-qat-GGUF:UD-Q4_K_XL"
+    ge4qatnt: "unsloth/gemma-4-E4B-it-qat-GGUF-nothink:UD-Q4_K_XL"
+    ge2qat: "unsloth/gemma-4-E2B-it-qat-GGUF:UD-Q4_K_XL"
+    ge2qatnt: "unsloth/gemma-4-E2B-it-qat-GGUF-nothink:UD-Q4_K_XL"
+    ge4mtp: "unsloth/gemma-4-E4B-it-qat-GGUF-mtp:UD-Q4_K_XL"
+    ge4mtpnt: "unsloth/gemma-4-E4B-it-qat-GGUF-mtp-nothink:UD-Q4_K_XL"
+    ge4ht: "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF:Q4_K_M"
+    ge4hnt: "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF-nothink:Q4_K_M"
+    ge4hmtp: "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF-mtp:Q4_K_M"
+    ge4hmnt: "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF-mtp-nothink:Q4_K_M"
     q36t: "unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL"
     q36nt: "unsloth/Qwen3.6-35B-A3B-GGUF-nothink:UD-Q4_K_XL"
     haut: "HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive:Q4_K_M"
@@ -54,10 +69,11 @@ matrix:
 
   sets:
     # any LLM can run alongside the small always-on model + STT + TTS (all CPU, no VRAM cost)
-    with_q8: "(coder | q35t | q35nt | q35ht | q35hnt | q4t | q4nt | q4ht | q4hnt | g26xl | g26q2 | ge4xl | ge2xl | q36t | q36nt | haut | haunt | mtpt | mtpnt) & q8 & stt"
+    with_q8: "(coder | q35t | q35nt | q35ht | q35hnt | q4t | q4nt | q4ht | q4hnt | g26xl | g26xlnt | g26mtp | g26mtpnt | g26ht | g26hnt | g26hmtp | g26hmnt | ge4qat | ge4qatnt | ge2qat | ge2qatnt | ge4mtp | ge4mtpnt | ge4ht | ge4hnt | ge4hmtp | ge4hmnt | q36t | q36nt | haut | haunt | mtpt | mtpnt) & q8 & stt"
     # FLUX runs alone — evicts everything including q8, but keeps STT for voice during image gen
     image_gen: "flux & stt"
 
+
 models:
   "Qwen3-Coder-Next-GGUF:Q4_K_M":
     cmd: |
@@ -151,38 +167,200 @@ models:
         ${qwen35_nothink_args}
         ${common_args}
 
-  "gemma-4-26B-A4B-it:UD-Q4_K_XL":
+  "gemma-4-26B-A4B-it-qat:UD-Q4_K_XL":
     cmd: |
       llama-server
-        -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL \
+        -hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL \
         ${ctx_256k}
         ${gemma4_sampling}
         ${common_args}
 
-  "gemma-4-26B-A4B-it:UD-Q2_K_XL":
+  "gemma-4-26B-A4B-it-qat-nothink:UD-Q4_K_XL":
     cmd: |
       llama-server
-        -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q2_K_XL \
+        -hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL \
+        ${ctx_256k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
+  "gemma-4-26B-A4B-it-qat-mtp:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL \
+        --spec-draft-hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q8_0-MTP \
+        --spec-type draft-mtp
+        --spec-draft-n-max 1
+        --swa-full
+        --kv-unified
+        --parallel 1
         ${ctx_256k}
         ${gemma4_sampling}
         ${common_args}
 
-  "unsloth/gemma-4-E4B-it-GGUF:UD-Q4_K_XL":
+  "gemma-4-26B-A4B-it-qat-mtp-nothink:UD-Q4_K_XL":
     cmd: |
       llama-server
-        -hf unsloth/gemma-4-E4B-it-GGUF:UD-Q4_K_XL \
+        -hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:UD-Q4_K_XL \
+        --spec-draft-hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q8_0-MTP \
+        --spec-type draft-mtp
+        --spec-draft-n-max 1
+        --swa-full
+        --kv-unified
+        --parallel 1
+        ${ctx_256k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
+  "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF:UD-Q4_K_XL \
+        ${ctx_256k}
+        ${gemma4_sampling}
+        ${common_args}
+
+  "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF-nothink:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF:UD-Q4_K_XL \
+        ${ctx_256k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
+  # The heretic QAT repo does not ship an MTP drafter,
+  # so borrow the one from the non-heretic unsloth QAT repo.
+  "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF-mtp:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF:UD-Q4_K_XL \
+        --spec-draft-hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q8_0-MTP \
+        --spec-type draft-mtp
+        --spec-draft-n-max 1
+        --swa-full
+        --kv-unified
+        --parallel 1
+        ${ctx_256k}
+        ${gemma4_sampling}
+        ${common_args}
+
+  "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF-mtp-nothink:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF:UD-Q4_K_XL \
+        --spec-draft-hf unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q8_0-MTP \
+        --spec-type draft-mtp
+        --spec-draft-n-max 1
+        --swa-full
+        --kv-unified
+        --parallel 1
+        ${ctx_256k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
+  "unsloth/gemma-4-E4B-it-qat-GGUF:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf unsloth/gemma-4-E4B-it-qat-GGUF:UD-Q4_K_XL \
         ${ctx_128k}
         ${gemma4_sampling}
         ${common_args}
 
-  "unsloth/gemma-4-E2B-it-GGUF:UD-Q4_K_XL":
+  "unsloth/gemma-4-E4B-it-qat-GGUF-nothink:UD-Q4_K_XL":
     cmd: |
       llama-server
-        -hf unsloth/gemma-4-E2B-it-GGUF:UD-Q4_K_XL \
+        -hf unsloth/gemma-4-E4B-it-qat-GGUF:UD-Q4_K_XL \
+        ${ctx_128k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
+  "unsloth/gemma-4-E2B-it-qat-GGUF:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf unsloth/gemma-4-E2B-it-qat-GGUF:UD-Q4_K_XL \
         ${ctx_128k}
         ${gemma4_sampling}
         ${common_args}
 
+  "unsloth/gemma-4-E2B-it-qat-GGUF-nothink:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf unsloth/gemma-4-E2B-it-qat-GGUF:UD-Q4_K_XL \
+        ${ctx_128k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
+  "unsloth/gemma-4-E4B-it-qat-GGUF-mtp:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf unsloth/gemma-4-E4B-it-qat-GGUF:UD-Q4_K_XL \
+        --spec-draft-hf unsloth/gemma-4-E4B-it-qat-GGUF:Q8_0-MTP \
+        --spec-type draft-mtp
+        --spec-draft-n-max 1
+        --swa-full
+        --kv-unified
+        --parallel 1
+        ${ctx_128k}
+        ${gemma4_sampling}
+        ${common_args}
+
+  "unsloth/gemma-4-E4B-it-qat-GGUF-mtp-nothink:UD-Q4_K_XL":
+    cmd: |
+      llama-server
+        -hf unsloth/gemma-4-E4B-it-qat-GGUF:UD-Q4_K_XL \
+        --spec-draft-hf unsloth/gemma-4-E4B-it-qat-GGUF:Q8_0-MTP \
+        --spec-type draft-mtp
+        --spec-draft-n-max 1
+        --swa-full
+        --kv-unified
+        --parallel 1
+        ${ctx_128k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
+  "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF:Q4_K_M":
+    cmd: |
+      llama-server
+        -hf llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF:Q4_K_M \
+        ${ctx_128k}
+        ${gemma4_sampling}
+        ${common_args}
+
+  "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF-nothink:Q4_K_M":
+    cmd: |
+      llama-server
+        -hf llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF:Q4_K_M \
+        ${ctx_128k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
+  "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF-mtp:Q4_K_M":
+    cmd: |
+      llama-server
+        -hf llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF:Q4_K_M \
+        --spec-draft-hf unsloth/gemma-4-E4B-it-qat-GGUF:Q8_0-MTP \
+        --spec-type draft-mtp
+        --spec-draft-n-max 1
+        --swa-full
+        --kv-unified
+        --parallel 1
+        ${ctx_128k}
+        ${gemma4_sampling}
+        ${common_args}
+
+  "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF-mtp-nothink:Q4_K_M":
+    cmd: |
+      llama-server
+        -hf llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF:Q4_K_M \
+        --spec-draft-hf unsloth/gemma-4-E4B-it-qat-GGUF:Q8_0-MTP \
+        --spec-type draft-mtp
+        --spec-draft-n-max 1
+        --swa-full
+        --kv-unified
+        --parallel 1
+        ${ctx_128k}
+        ${gemma4_nothink_sampling}
+        ${common_args}
+
   "unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL":
     cmd: |
       llama-server
@@ -235,20 +413,14 @@ models:
       --parallel 1
       ${common_args}
 
-  # STT via whisper.cpp (Vulkan GPU on RX 580, always loaded, ~600MB VRAM)
-  # Model auto-downloaded by init container, see deployment.yaml
-  # Note: Vulkan whisper on AMD GPUs has known quality issues on some cards;
-  # if transcriptions come out as garbage/gibberish, add --no-gpu to fall back.
-  "whisper-small":
+  # STT via parakeet-server (parakeet.cpp OpenAI-compatible server, CPU, always loaded)
+  # Model downloaded on first start and cached under /root/.cache/parakeet.cpp/models
+  # parakeet-proxy.py sits in front to convert any audio format to WAV via ffmpeg,
+  # since parakeet-server only accepts real WAV but browsers send Ogg/Opus.
+  "parakeet-tdt_ctc-1.1b":
     checkEndpoint: none
     cmd: |
-      whisper-server
-        --port ${PORT}
-        -m /root/.cache/whisper/ggml-small.bin
-        --request-path /v1/audio
-        --inference-path /transcriptions
-        --convert
-        --threads 6
+      env PROXY_PORT=${PORT} FFMPEG_BIN=/root/.cache/ffmpeg/ffmpeg python3 /config/parakeet-proxy.py
 
 
   # Image generation via stable-diffusion.cpp (sd-server)

apps/llama/configs/parakeet-proxy.py

@@ -0,0 +1,227 @@
+#!/usr/bin/env python3
+"""
+Thin reverse proxy for parakeet-server.
+
+Accepts POST /v1/audio/transcriptions with any audio format,
+converts the audio to 16 kHz mono WAV via ffmpeg, then forwards
+the converted file to the real parakeet-server running on PARAKEET_PORT.
+
+Also proxies GET /health straight through.
+
+Usage:
+    PROXY_PORT=<port>  PARAKEET_PORT=<upstream>  python3 parakeet-proxy.py
+"""
+
+import http.server
+import io
+import os
+import subprocess
+import sys
+import tempfile
+import urllib.request
+import urllib.error
+
+PROXY_PORT    = int(os.environ.get("PROXY_PORT", "8080"))
+PARAKEET_PORT = PROXY_PORT + 1
+FFMPEG        = os.environ.get("FFMPEG_BIN", "ffmpeg")
+MODEL         = os.environ.get("PARAKEET_MODEL", "tdt_ctc-1.1b-q4_k.gguf")
+CACHE_DIR     = os.environ.get("PARAKEET_CACHE_DIR", "/root/.cache/parakeet.cpp/models")
+
+
+def convert_to_wav(data: bytes) -> bytes:
+    """Convert any audio bytes to 16 kHz mono PCM WAV via ffmpeg."""
+    with tempfile.NamedTemporaryFile(suffix=".input", delete=False) as inf:
+        inf.write(data)
+        inf_path = inf.name
+    out_path = inf_path + ".wav"
+    try:
+        subprocess.run(
+            [
+                FFMPEG, "-y",
+                "-i", inf_path,
+                "-ar", "16000",
+                "-ac", "1",
+                "-f", "wav",
+                out_path,
+            ],
+            check=True,
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.DEVNULL,
+        )
+        with open(out_path, "rb") as f:
+            return f.read()
+    finally:
+        os.unlink(inf_path)
+        if os.path.exists(out_path):
+            os.unlink(out_path)
+
+
+def parse_multipart(content_type: str, body: bytes):
+    """
+    Parse a multipart/form-data body.
+    Returns a dict of field_name -> (filename_or_None, content_type, data).
+    """
+    import email
+    from email import policy as email_policy
+
+    # email.parser needs the full MIME headers to parse multipart
+    raw = b"Content-Type: " + content_type.encode() + b"\r\n\r\n" + body
+    msg = email.message_from_bytes(raw, policy=email_policy.compat32)
+    parts = {}
+    for part in msg.get_payload():
+        cd = part.get("Content-Disposition", "")
+        name = None
+        filename = None
+        for item in cd.split(";"):
+            item = item.strip()
+            if item.startswith('name='):
+                name = item[5:].strip('"')
+            elif item.startswith('filename='):
+                filename = item[9:].strip('"')
+        if name is not None:
+            parts[name] = (filename, part.get_content_type(), part.get_payload(decode=True))
+    return parts
+
+
+def build_multipart(fields: dict) -> tuple[bytes, str]:
+    """
+    Build a multipart/form-data body from fields dict:
+      field_name -> (filename_or_None, content_type, data_bytes)
+    Returns (body_bytes, content_type_header_value).
+    """
+    boundary = b"----ParakeetProxyBoundary0xDEADBEEF"
+    body = b""
+    for name, (filename, ct, data) in fields.items():
+        body += b"--" + boundary + b"\r\n"
+        if filename:
+            body += (
+                f'Content-Disposition: form-data; name="{name}"; filename="{filename}"\r\n'
+            ).encode()
+        else:
+            body += f'Content-Disposition: form-data; name="{name}"\r\n'.encode()
+        body += f"Content-Type: {ct}\r\n\r\n".encode()
+        body += data + b"\r\n"
+    body += b"--" + boundary + b"--\r\n"
+    return body, f"multipart/form-data; boundary={boundary.decode()}"
+
+
+class ProxyHandler(http.server.BaseHTTPRequestHandler):
+    def log_message(self, fmt, *args):
+        print(f"[parakeet-proxy] {self.address_string()} - {fmt % args}", flush=True)
+
+    def do_GET(self):
+        if self.path == "/health":
+            self._forward_get("/health")
+        else:
+            self.send_response(404)
+            self.end_headers()
+
+    def do_POST(self):
+        if self.path.rstrip("/") == "/v1/audio/transcriptions":
+            self._handle_transcription()
+        else:
+            self.send_response(404)
+            self.end_headers()
+
+    def _forward_get(self, path):
+        try:
+            url = f"http://127.0.0.1:{PARAKEET_PORT}{path}"
+            with urllib.request.urlopen(url, timeout=5) as resp:
+                body = resp.read()
+                self.send_response(resp.status)
+                self.send_header("Content-Type", resp.headers.get("Content-Type", "application/json"))
+                self.end_headers()
+                self.wfile.write(body)
+        except Exception as e:
+            self.send_response(502)
+            self.end_headers()
+            self.wfile.write(str(e).encode())
+
+    def _handle_transcription(self):
+        length = int(self.headers.get("Content-Length", 0))
+        body = self.rfile.read(length)
+        ct = self.headers.get("Content-Type", "")
+
+        try:
+            fields = parse_multipart(ct, body)
+        except Exception as e:
+            self._error(400, f"failed to parse multipart: {e}")
+            return
+
+        if "file" not in fields:
+            self._error(400, "missing required field 'file'")
+            return
+
+        filename, file_ct, audio_data = fields["file"]
+
+        # Convert to WAV regardless of what we received
+        try:
+            wav_data = convert_to_wav(audio_data)
+        except subprocess.CalledProcessError:
+            self._error(400, "ffmpeg could not decode audio")
+            return
+        except Exception as e:
+            self._error(500, f"conversion error: {e}")
+            return
+
+        # Rebuild multipart with converted WAV, preserve other fields
+        new_fields = {}
+        for name, (fn, fct, data) in fields.items():
+            if name == "file":
+                new_fields[name] = ("recording.wav", "audio/wav", wav_data)
+            else:
+                new_fields[name] = (fn, fct, data)
+
+        new_body, new_ct = build_multipart(new_fields)
+
+        # Forward to parakeet-server
+        try:
+            url = f"http://127.0.0.1:{PARAKEET_PORT}/v1/audio/transcriptions"
+            req = urllib.request.Request(
+                url,
+                data=new_body,
+                headers={"Content-Type": new_ct},
+                method="POST",
+            )
+            with urllib.request.urlopen(req, timeout=300) as resp:
+                resp_body = resp.read()
+                self.send_response(resp.status)
+                self.send_header("Content-Type", resp.headers.get("Content-Type", "application/json"))
+                self.end_headers()
+                self.wfile.write(resp_body)
+        except urllib.error.HTTPError as e:
+            resp_body = e.read()
+            self.send_response(e.code)
+            self.send_header("Content-Type", e.headers.get("Content-Type", "application/json"))
+            self.end_headers()
+            self.wfile.write(resp_body)
+        except Exception as e:
+            self._error(502, f"upstream error: {e}")
+
+    def _error(self, code: int, msg: str):
+        body = f'{{"error":{{"message":"{msg}","type":"proxy_error"}}}}'.encode()
+        self.send_response(code)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+        self.wfile.write(body)
+
+
+if __name__ == "__main__":
+    proc = subprocess.Popen([
+        "parakeet-server",
+        "--host", "127.0.0.1",
+        "--port", str(PARAKEET_PORT),
+        "--model", MODEL,
+        "--cache-dir", CACHE_DIR,
+    ])
+    print(f"[parakeet-proxy] started parakeet-server pid={proc.pid} on :{PARAKEET_PORT}", flush=True)
+
+    server = http.server.HTTPServer(("0.0.0.0", PROXY_PORT), ProxyHandler)
+    print(f"[parakeet-proxy] listening on :{PROXY_PORT}", flush=True)
+    try:
+        server.serve_forever()
+    except KeyboardInterrupt:
+        pass
+    finally:
+        proc.terminate()
+        proc.wait()

apps/llama/deployment.yaml

@@ -18,7 +18,7 @@ spec:
     spec:
       initContainers:
         - name: download-whisper
-          image: ghcr.io/mostlygeek/llama-swap:unified-vulkan-2026-06-03
+          image: gitea.lumpiasty.xyz/lumpiasty/llama-swap:unified-vulkan-parakeet-2026-06-12
           command:
             - sh
             - -c
@@ -48,7 +48,7 @@ spec:
               mountPath: /root/.cache
       containers:
         - name: llama-swap
-          image: ghcr.io/mostlygeek/llama-swap:unified-vulkan-2026-06-03
+          image: gitea.lumpiasty.xyz/lumpiasty/llama-swap:unified-vulkan-parakeet-2026-06-12
           imagePullPolicy: IfNotPresent
           command:
             - llama-swap

apps/llama/kokoro.yaml

@@ -21,7 +21,7 @@ spec:
           # OpenAI-compatible Kokoro-FastAPI TTS server, CPU PyTorch backend.
           # Models baked into the image (no PVC needed).
           # v0.3.0 includes fix for per-request voice tensor memory leak (#459).
-          image: ghcr.io/remsky/kokoro-fastapi-cpu:v0.4.0
+          image: ghcr.io/remsky/kokoro-fastapi-cpu:v0.5.0
           ports:
             - containerPort: 8880
               name: http

apps/llama/kustomization.yaml

@@ -13,3 +13,4 @@ configMapGenerator:
     namespace: llama
     files:
       - config.yaml=configs/config.yaml
+      - parakeet-proxy.py=configs/parakeet-proxy.py

apps/nas/deployment.yaml

@@ -15,7 +15,7 @@ spec:
     spec:
       initContainers:
         - name: prepare-home
-          image: alpine:3.23.4
+          image: alpine:3.24.1
           imagePullPolicy: IfNotPresent
           command:
             - /bin/sh

apps/openwebui/release.yaml

@@ -74,11 +74,17 @@ spec:
         value: "false"
       - name: OAUTH_MERGE_ACCOUNTS_BY_EMAIL
         value: "true"
+      - name: ENABLE_OAUTH_ROLE_MANAGEMENT
+        value: "true"
+      - name: OAUTH_ROLES_CLAIM
+        value: "groups"
+      - name: OAUTH_ADMIN_ROLES
+        value: "Admins"
       - name: OAUTH_ALLOWED_ROLES
-        value: "user"
+        value: "Users"
       - name: OAUTH_AUTO_REDIRECT
         value: "true"
-      # STT via whisper-server (routed through llama-swap)
+      # STT via parakeet-server (routed through llama-swap)
       - name: AUDIO_STT_ENGINE
         value: "openai"
       - name: AUDIO_STT_OPENAI_API_BASE_URL
@@ -86,9 +92,7 @@ spec:
       - name: AUDIO_STT_OPENAI_API_KEY
         value: "ignored"
       - name: AUDIO_STT_MODEL
-        value: "whisper-small"
-      - name: AUDIO_STT_SUPPORTED_CONTENT_TYPES
-        value: "audio/wav,audio/wave"
+        value: "parakeet-tdt_ctc-1.1b"
       # TTS via OuteTTS (routed through llama-swap)
       # TTS via dedicated Kokoro server (CPU-only, separate pod)
       - name: AUDIO_TTS_ENGINE

docs/coredns.md

@@ -0,0 +1,110 @@
+# CoreDNS resolver
+
+## Goal
+
+Replace the RouterOS built-in DNS forwarder with a CoreDNS container for
+configurability, and suppress IPv6 (AAAA) resolution by default to keep traffic
+on IPv4.
+
+## Background
+
+The ISP provides no native IPv6 — only a Hurricane Electric (HE) tunnel
+(`2001:470:61a3::/48`). HE addresses fall in ranges some sites flag as
+datacenter/bot traffic, producing endless CAPTCHAs. The goal is to prefer IPv4
+egress while keeping IPv6 available for our own services and any domain
+explicitly trusted over IPv6.
+
+## What this is NOT (and why)
+
+An earlier iteration used **DNS64 + NAT64 (Tayga)** to force traffic through
+IPv4. It was removed:
+
+- **Performance**: Tayga is a userspace translator with no hardware offload.
+  Every translated packet crossed RouterOS twice (v6 in, v4 out) plus a
+  userspace hop, capping throughput at ~250 Mbps against a 1 Gbps line.
+- **SPOF**: two containers (CoreDNS + Tayga) in the datapath of nearly all
+  traffic on a router whose native forwarder had been rock-solid.
+- **Architectural inversion**: NAT64 exists to let IPv6-only clients reach IPv4.
+  We don't want IPv6 egress at all — using NAT64 to avoid IPv6 was solving the
+  problem backwards.
+
+Plain AAAA suppression in CoreDNS achieves the same IPv4-preferred outcome with
+zero datapath overhead — DNS is the only thing touched, packet forwarding stays
+on the RouterOS fastpath at line rate.
+
+The full account of the NAT64/IPv6-mostly attempt and why it was abandoned is in
+[nat64-dns64-postmortem.md](./nat64-dns64-postmortem.md).
+
+## How it works
+
+CoreDNS runs as a single container (`172.20.0.3`), reachable from RouterOS DNS
+which forwards client queries to it. The [Corefile](../mikrotik/coredns/Corefile)
+has three server blocks:
+
+1. **`lumpiasty.xyz`** — our own zone. Forwards normally, keeps AAAA, so internal
+   services reachable over the HE prefix resolve to their real IPv6 addresses.
+2. **`.` (default)** — forwards everything else, but a `template IN AAAA` block
+   returns empty NOERROR for all AAAA queries, so clients fall back to IPv4 and
+   avoid the HE tunnel's flagged egress. A records and all other types pass
+   through untouched.
+
+The whitelist is implemented as a reusable `(aaaa_allowed)` snippet imported by
+zones that should keep AAAA. To trust another domain over IPv6, add a server
+block for it that imports `aaaa_allowed`.
+
+### Why suppression, not NXDOMAIN
+
+The AAAA template returns NOERROR with an empty answer (NODATA), not NXDOMAIN.
+This is correct: the name exists, it just has no (advertised) AAAA. Clients
+treat it as "no IPv6 address" and use the A record. Returning NXDOMAIN would
+wrongly imply the name doesn't exist and break the A lookup.
+
+## Future improvement
+
+The current global-suppress-plus-whitelist is coarse: a domain that is genuinely
+IPv6-only (no A record) and not whitelisted becomes unreachable. The intended
+end state is a plugin that suppresses AAAA only when the domain also has an A
+record, so IPv6-only destinations keep working without manual whitelisting. No
+in-tree CoreDNS plugin does this today.
+
+## Custom image
+
+Built from source with a minimal plugin set (`errors`, `log`, `health`,
+`template`, `cache`, `forward`, `reload`) instead of the default ~40, producing
+a ~6-8 MB image. The `dns64` plugin is no longer compiled in.
+
+Source: [`mikrotik/coredns/`](../mikrotik/coredns/). Built by Woodpecker
+([`.woodpecker/coredns-build.yaml`](../.woodpecker/coredns-build.yaml)) on pushes
+touching `mikrotik/coredns/**`, pushed to `gitea.lumpiasty.xyz/lumpiasty/coredns-mikrotik`.
+
+## RouterOS integration
+
+- `/ip/dns servers=172.20.0.3` — RouterOS forwards client queries to CoreDNS
+- RDNSS in RA (`/ipv6/nd dns=...` on vlan2/vlan5) advertises an IPv6 resolver
+  (the router's per-VLAN address) to dual-stack clients; RouterOS DNS relays to
+  CoreDNS
+- No DHCP option 108, no PREF64 — those belonged to the removed IPv6-mostly setup
+
+## Pitfalls learned (kept for reference)
+
+These were hit during the NAT64 era and the migration; some still apply:
+
+1. **RouterOS static FWD entries corrupt NXDOMAIN.** A `type=FWD match-subdomain=yes`
+   entry returns NOERROR/empty instead of relaying NXDOMAIN. Combined with
+   `ndots:5` and kubernetes pod search domains, `getaddrinfo` stops at the first
+   search-suffixed NODATA candidate and never tries the absolute name — apps fail
+   with `ENOTFOUND` for valid hostnames while `nslookup` (absolute query) works.
+   Our own zone is therefore handled in the Corefile, not via a RouterOS FWD
+   entry. RouterOS DNS does plain forwarding only (plus the Tailscale `ts.net`
+   FWD, which is acceptable as its subdomains genuinely don't exist publicly).
+2. **`advertise-dns=no` on new ND entries.** RouterOS creates per-interface
+   `ipv6 nd` entries with `advertise-dns=no`, suppressing the RDNSS option even
+   when a static `dns=` list is set. Must be enabled explicitly.
+3. **Per-interface ND entries must be created, not modified.** Only the
+   `interface=all` default ships out of the box; `api_find_and_modify` matching a
+   specific interface silently matches nothing. Use `api_modify`.
+
+Verification: `rdisc6` (NixOS package `ndisc6`) dumps RA contents. The CoreDNS
+`log` plugin output is visible via `/log print` on the router (container
+`logging=yes`) and shows the rcode CoreDNS returned — comparing it to what the
+client received isolates which hop corrupts a response.

docs/lte-failover-design.md

@@ -84,9 +84,10 @@ subnets would fail routing lookup with "net unreachable" without it.
 
 | Destination | Source | Distance | Active when |
 |---|---|---|---|
-| `0.0.0.0/0` | static via `pppoe-gpon` | 1 | GPON up |
+| `1.0.0.1/32`, `8.8.4.4/32` | static via `pppoe-gpon` | 1 | always |
+| `0.0.0.0/0` | static via `1.0.0.1`, `8.8.4.4` (recursive) | 1, 2 | GPON ping check succeeds |
 | `0.0.0.0/0` | BGP from D-Link via `192.168.6.2` | 200 | wwan up on D-Link |
-| `2000::/3` | static via `sit1` (HE tunnel) | 1 | sit1 active (HE tunnel works) |
+| `2000::/3` | static via `2001:470:70:dd::1` (HE tunnel) | 1 | HE tunnel ping check succeeds |
 | `2000::/3` | BGP from D-Link via `2001:470:61a3:600::2` | 200 | wwan up on D-Link |
 
 RouterOS distance comparison is straightforward: distance 1 always wins
@@ -136,11 +137,12 @@ preferred route for D-Link's own traffic.
 - **wwan modem goes down** → BIRD2 device protocol detects wwan0 down →
   static `lte_default` / `lte_default6` become unreachable → BGP withdraws
   announcements → CRS removes BGP-learned default
-- **GPON drops** → `pppoe-gpon` interface down → CRS distance-1 default
-  route inactive → distance-200 BGP route activates → CRS withdraws its
-  default-originate announcement to D-Link (since no default is installed
-  any more) → D-Link's kernel default-via-CRS is removed → D-Link uses
-  wwan kernel default → traffic flows from CRS via vlan6 → D-Link → wwan
+- **GPON drops or blackholes** → recursive ping checks (1.0.0.1, 8.8.4.4) over `pppoe-gpon`
+  fail (takes ~20s: 10s ping interval + 10s timeout) → CRS distance-1/2 default routes inactive → distance-200 BGP route
+  activates → CRS withdraws its default-originate announcement to D-Link (loop
+  prevention prevents reflecting D-Link's own route) → D-Link's kernel
+  default-via-CRS is removed → D-Link uses wwan kernel default → traffic flows
+  from CRS via vlan6 → D-Link → wwan
 
 All transitions are automatic and driven by interface state. No active
 probing (Netwatch / mwan3), no scripts toggling routes.
@@ -241,6 +243,16 @@ QMI initialization within ~1 second.
 
 Full investigation: see [wwan-bm806c-qmi-workaround.md](./wwan-bm806c-qmi-workaround.md).
 
+## Multi-WAN Stale Connection Tracking
+
+When the routing table fails over from GPON to LTE (or vice versa), RouterOS does not automatically clear existing connection tracking entries. If an established TCP/UDP connection is routed out the new WAN interface, it retains the NAT translation state (source IP) of the old WAN interface. The packet is sent to the ISP with the wrong source IP and is silently dropped, causing clients (like Tailscale) to hang for minutes until their internal sockets time out.
+
+To solve this purely declaratively without scripts or blanket connection flushes, the `forward` chain is configured to "fast-fail" these shifted connections:
+
+1. Connections are marked with their egress WAN upon establishment (`wan-gpon` or `wan-lte`) via the `mangle` table.
+2. If an established connection with a `wan-gpon` mark attempts to route out `vlan6` (LTE), or a `wan-lte` mark routes out `pppoe-gpon`, it is explicitly rejected (`tcp-reset` for TCP, `icmp-network-unreachable` for UDP) before reaching the NAT table.
+3. This rejection immediately signals the client OS that the route is dead, forcing the application (Tailscale, SIP clients, etc.) to instantly close the socket and establish a new one, which successfully binds to the new WAN interface and NAT state.
+
 ## Implementation files
 
 | File | Role |

docs/nat64-dns64-postmortem.md

@@ -0,0 +1,136 @@
+# Postmortem: NAT64 / IPv6-mostly attempt
+
+A record of an architecture that was built, run for ~2 days, and removed. Kept
+so the reasoning isn't re-discovered the hard way. For the current DNS setup see
+[coredns.md](./coredns.md); for network overview see [network.md](./network.md).
+
+## The original problem
+
+The ISP provides no native IPv6 — only a Hurricane Electric (HE) 6in4 tunnel
+(`2001:470:61a3::/48`). HE address ranges are widely classified as
+datacenter/hosting space, so some sites (Google, Cloudflare-fronted services,
+various login flows) treat IPv6 traffic from them as bot/VPN traffic: endless
+CAPTCHAs, "unusual traffic" interstitials, or outright blocks. IPv4 egress
+(the ISP's residential PPPoE address) is unaffected.
+
+The goal: keep using the network normally without IPv6 triggering these flags,
+while still wanting some IPv6 (e.g. inbound to self-hosted services).
+
+## What was built
+
+An **IPv6-mostly** network (RFC 8925) with **DNS64 + NAT64**, intended to push
+egress onto IPv4 while presenting IPv6 to clients:
+
+- **CoreDNS container** with the `dns64` plugin (`translate_all`): synthesized
+  `64:ff9b::/96` AAAA records from A records for *all* names, so even dual-stack
+  destinations resolved to a NAT64 address.
+- **Tayga container** (`ghcr.io/apalrd/tayga-nat64`): stateless NAT64 translator.
+  IPv6 traffic to `64:ff9b::/96` was routed to it, translated to IPv4, and
+  masqueraded out the GPON PPPoE interface. So all "IPv6" egress actually left
+  as IPv4 on the residential address — bypassing the HE tunnel and its flagging.
+- **RouterOS RA + DHCP**: DHCP option 108 (IPv6-only preferred) to make capable
+  clients drop IPv4, PREF64 (RFC 8781) to advertise the NAT64 prefix for CLAT,
+  RDNSS (RFC 8106) to hand IPv6-only clients a resolver.
+- Dedicated `nat64` bridge, `fc64::/126` link, `192.168.240.0/20` Tayga pool,
+  static routes, and firewall rules (including NAT64-mapped RFC1918 blocks to
+  prevent the translator being used as a policy bypass).
+
+## Why it was removed
+
+### 1. Performance — the dealbreaker
+
+Throughput collapsed from line rate (~1 Gbps) to **~200-300 Mbps**, saturating
+the router CPU. Causes, all structural:
+
+- Tayga is a **userspace** translator. Every translated packet leaves the kernel
+  fastpath, is copied to userspace, translated, and re-injected.
+- Translated traffic crosses RouterOS **twice** — once as IPv6 (LAN → Tayga),
+  once as IPv4 (Tayga → WAN, with masquerade) — doubling firewall/conntrack work.
+- No hardware offload or fasttrack applies to either leg.
+
+With `translate_all`, *nearly all* internet traffic went through this path, so
+the penalty hit everything, not just IPv4-only destinations.
+
+### 2. Single point of failure
+
+DNS (CoreDNS) and most of the datapath (Tayga) became two containers in the
+critical path on a router whose built-in forwarder had been completely reliable.
+Container restarts, image pulls, or a crash now took down connectivity.
+
+### 3. Architectural inversion
+
+NAT64 exists to let **IPv6-only** clients reach the **IPv4** internet. The actual
+goal here was the opposite — *avoid* IPv6 egress entirely. Building an IPv6-only
+client environment (option 108, CLAT, PREF64) and then translating all of it back
+to IPv4 was solving the problem backwards. The complexity existed only to route
+around a property of the HE tunnel.
+
+### 4. Firewall complexity and a translation bypass hole
+
+NAT64 punched a hole in the firewall model. RouterOS filters IPv4 and IPv6
+independently, but NAT64 traffic enters as IPv6 and *leaves* as IPv4 after
+translation — so the carefully-built IPv4 forward policy (inter-VLAN isolation,
+RFC1918-to-WAN blocks) was simply bypassed for anything arriving via the
+translator. A client could reach a private IPv4 range by encoding it in the
+NAT64 prefix (`64:ff9b::c0a8:xxyy` = `192.168.x.y`), and the IPv4 rules would
+never see it because the packet was IPv6 until Tayga rewrote it.
+
+Plugging this required mirroring the IPv4 policy in the IPv6 chain: explicit
+`reject` rules for every NAT64-mapped RFC1918 block (`64:ff9b::a00:0/104`,
+`64:ff9b::ac10:0/108`, `64:ff9b::c0a8:0/112`), per-VLAN accept rules toward the
+`nat64` interface, plus a separate masquerade and LB hairpin-accept for the
+Tayga pool. That is a parallel, easy-to-get-wrong copy of the existing ruleset,
+whose correctness depended on getting CIDR-to-prefix arithmetic right. Removing
+NAT64 deleted all of it.
+
+### 5. Operational fragility (see coredns.md for detail)
+
+The setup had a long tail of subtle failure modes, each presenting identically
+as "client can't connect":
+
+- RouterOS static `FWD` entries return `NOERROR`/empty instead of relaying
+  `NXDOMAIN`, which broke `getaddrinfo` search-domain handling in Kubernetes
+  pods (`ENOTFOUND` for valid names).
+- `translate_all` discarded real AAAA for IPv6-only internal services, and
+  returned empty answers for names with no A record.
+- Per-interface RouterOS `ipv6 nd` entries default to `advertise-dns=no` and must
+  be *created* (not modified), so RDNSS/PREF64 silently never advertised.
+- Dynamic `from-pool` VLAN addressing made advertised RDNSS addresses point at
+  nonexistent router addresses.
+- Option 108 honoured by clients before the NAT64 path was verified working left
+  them stuck "obtaining IP address".
+
+Each was individually fixable, but the aggregate was a brittle system whose
+benefit didn't justify the surface area.
+
+## What replaced it
+
+Plain CoreDNS forwarder with **AAAA suppression by default** plus a whitelist for
+domains that should keep IPv6 (our own zone over the HE prefix, and any explicitly
+trusted domain). Clients prefer IPv4 because they simply don't receive AAAA for
+most names — no translation, no extra datapath hop, packet forwarding stays on the
+RouterOS fastpath at line rate. DNS is the only thing in the path. See
+[coredns.md](./coredns.md).
+
+Tradeoff accepted: a non-whitelisted IPv6-only destination (no A record) is
+unreachable. In practice essentially everything on the public internet still has
+an A record. The intended future refinement is a CoreDNS plugin that suppresses
+AAAA only when an A record also exists, removing the need for the whitelist; no
+in-tree plugin does this today.
+
+## Lessons
+
+- **Measure throughput before committing to an in-path translator on SOHO-class
+  hardware.** Userspace NAT64 (Tayga/Jool-in-container) on a MikroTik CPU is
+  fine for a few hundred Mbps, not for saturating a gigabit line.
+- **Match the mechanism to the actual goal.** The goal was "prefer IPv4 egress",
+  which is a one-line DNS policy, not a transition technology.
+- **Prefer solutions that stay on the fastpath.** Anything that pulls bulk
+  traffic into userspace or doubles the forwarding work will dominate the CPU.
+- **Fewer moving parts in the critical path.** Two containers carrying all DNS
+  and most traffic is a worse availability story than the stock forwarder, for a
+  cosmetic benefit (avoiding CAPTCHAs on some sites).
+- **Protocol translation breaks the firewall model.** When traffic changes L3
+  protocol mid-path, the two firewall policies must be kept in sync by hand, and
+  any gap is a silent bypass. A solution that doesn't translate keeps a single
+  coherent policy.

docs/network.md

@@ -93,8 +93,8 @@ There are also networks, which are not VLANs, but are routed:
   Static assignment on CRS, access to factory IP of ONU
 - Containers on CRS<br>
   Access to every other network<br>
-  IP: 172.17.0.1/16, 2001:470:61a3:500::/64<br>
-  Static IP management
+  IP: 172.20.0.1/24, 2001:470:61a3:500::/64<br>
+  Static IP management, hosts Tailscale and CoreDNS containers
 
 Whole network is designed to eliminate VLANs, overlays where unnecessary to keep things simple. Only NAT rules are:
 
@@ -105,6 +105,10 @@ Whole network is designed to eliminate VLANs, overlays where unnecessary to keep
   Tailscale assigns IPv6 from private subnet with no way to configure it, so the assigned IPs are not routable
 - IPv4 port forwards from GPON PPPoE to respective services
 
+## DNS and IPv6 preference
+
+DNS is served by a CoreDNS container (`172.20.0.3`); RouterOS forwards client queries to it. CoreDNS suppresses AAAA records by default so clients prefer IPv4, avoiding the HE tunnel's datacenter-flagged egress (which triggers CAPTCHAs on some sites). Our own zone (`lumpiasty.xyz`) and any explicitly whitelisted domains keep AAAA for native IPv6. See [CoreDNS resolver](./coredns.md). An earlier NAT64/IPv6-mostly approach to the same problem was built and abandoned; see the [postmortem](./nat64-dns64-postmortem.md).
+
 There is also an UPnP and NAT-PMP enabled to automatically configure port forwards from LAN.
 
 ## Uplink

infra/controllers/cert-manager-webhook-ovh.yaml

@@ -18,7 +18,7 @@ spec:
   chart:
     spec:
       chart: cert-manager-webhook-ovh
-      version: 0.9.10
+      version: 0.9.13
       sourceRef:
         kind: HelmRepository
         name: cert-manager-webhook-ovh

infra/controllers/cilium.yaml

@@ -23,7 +23,7 @@ spec:
   chart:
     spec:
       chart: cilium
-      version: 1.19.4
+      version: 1.19.5
       sourceRef:
         kind: HelmRepository
         name: cilium

infra/controllers/cloudnative-pg.yaml

@@ -23,7 +23,7 @@ spec:
   chart:
     spec:
       chart: cloudnative-pg
-      version: 0.28.2
+      version: 0.28.3
       sourceRef:
         kind: HelmRepository
         name: cnpg

infra/controllers/openbao.yaml

@@ -23,7 +23,7 @@ spec:
   chart:
     spec:
       chart: openbao
-      version: 0.28.3
+      version: 0.28.4
       sourceRef:
         kind: HelmRepository
         name: openbao

infra/controllers/openebs.yaml

@@ -23,7 +23,7 @@ spec:
   chart:
     spec:
       chart: openebs
-      version: 4.4.0
+      version: 4.5.1
       sourceRef:
         kind: HelmRepository
         name: openebs

mikrotik/README.md

@@ -0,0 +1,43 @@
+# Mikrotik containers
+
+RouterOS containers running on the CRS418 providing network services that
+RouterOS cannot handle natively.
+
+## CoreDNS
+
+Replaces the built-in RouterOS DNS forwarder. Plain forwarding resolver with
+selective AAAA suppression: AAAA is suppressed by default so clients prefer IPv4
+(avoiding the HE tunnel's datacenter-flagged egress), while our own zone and any
+whitelisted domains keep AAAA for native IPv6.
+
+Source: [`coredns/`](coredns/). Image built by Woodpecker CI
+([`.woodpecker/coredns-build.yaml`](../.woodpecker/coredns-build.yaml)), pushed to
+`gitea.lumpiasty.xyz/lumpiasty/coredns-mikrotik`.
+
+The Corefile is baked into the image — edit [`coredns/Corefile`](coredns/Corefile)
+and push; the pipeline rebuilds and pushes a new image. Custom-built with a
+minimal plugin set (~6-8 MB vs the official ~20 MB image) to fit the CRS flash.
+
+See [docs/coredns.md](../docs/coredns.md) for design rationale, including why
+the earlier NAT64/DNS64 approach was removed.
+
+### Why not the official coredns/coredns image?
+
+The official image ships ~40 plugins and weighs ~20 MB compressed. A custom build with the 7 plugins we actually need fits in ~6-8 MB — important for the CRS internal flash.
+
+## Deployment
+
+The router configuration (container definitions, veth interfaces, bridge ports,
+DNS settings, firewall) is managed declaratively via Ansible, not by manual CLI
+commands. See [`ansible/roles/routeros/`](../ansible/roles/routeros/) and run:
+
+```sh
+cd ansible && ansible-playbook playbooks/routeros.yml
+```
+
+Containers do not auto-start on first image pull; after the initial deploy,
+start manually once (subsequent boots are handled by `start-on-boot=yes`):
+
+```
+/container/start [find name=coredns]
+```

mikrotik/coredns/Corefile

@@ -0,0 +1,53 @@
+# CoreDNS as a plain forwarding resolver with selective AAAA suppression.
+#
+# Background: the ISP provides no native IPv6, only a Hurricane Electric tunnel.
+# HE addresses are flagged as datacenter ranges by some sites (endless CAPTCHAs,
+# bot detection). To avoid this, IPv6 (AAAA) resolution is suppressed by default
+# so clients use IPv4, while a whitelist keeps AAAA for domains where native
+# IPv6 is wanted (our own services reachable over the HE prefix, and any domain
+# explicitly trusted over IPv6).
+#
+# NAT64/DNS64 was tried and removed: it forced most traffic through a userspace
+# Tayga translator, capping throughput at ~250 Mbps on the RB-class CPU (line
+# rate is 1 Gbps) and adding two containers as a SPOF — all to avoid IPv6 egress
+# we don't want in the first place. Plain AAAA suppression achieves the same
+# IPv4-preferred outcome with zero datapath overhead.
+#
+# TODO: replace the global template suppression + whitelist with a plugin that
+# suppresses AAAA only when the domain has no A record (so IPv6-only
+# destinations still work). No such in-tree plugin exists yet.
+
+# Whitelist: domains that keep AAAA resolution (native IPv6 via HE tunnel).
+(aaaa_allowed) {
+    forward . 1.1.1.1 8.8.8.8 {
+        prefer_udp
+    }
+    cache 300
+    errors
+    log . {
+        class error
+    }
+}
+
+# Our own zone: services have native IPv6 on the HE prefix, keep AAAA.
+lumpiasty.xyz:53 {
+    import aaaa_allowed
+}
+
+# Default: forward everything, but suppress AAAA so clients use IPv4 and
+# avoid the HE tunnel's datacenter-flagged egress.
+.:53 {
+    template IN AAAA {
+        rcode NOERROR
+    }
+    forward . 1.1.1.1 8.8.8.8 {
+        prefer_udp
+    }
+    cache 300
+    errors
+    log . {
+        class error
+    }
+    reload
+    health :8080
+}

mikrotik/coredns/Dockerfile

@@ -0,0 +1,32 @@
+# Stage 1: build CoreDNS with minimal plugin set
+FROM golang:1.26-alpine AS build
+
+RUN apk add --no-cache git make bash
+
+WORKDIR /src
+RUN git clone --depth 1 --branch v1.12.1 \
+    https://github.com/coredns/coredns .
+
+# Overwrite plugin.cfg with our trimmed list before compilation
+COPY plugin.cfg .
+
+RUN go generate && make
+
+# Stage 2: extract CA certificates from a full image
+FROM debian:stable-slim AS certs
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ca-certificates && rm -rf /var/lib/apt/lists/*
+
+# Stage 3: minimal runtime — scratch + binary + certs only
+FROM scratch
+
+COPY --from=certs /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/
+COPY --from=build /src/coredns /coredns
+COPY Corefile /Corefile
+
+# 53:  DNS (UDP + TCP)
+# 8080: health endpoint
+EXPOSE 53/udp 53/tcp 8080/tcp
+
+# RouterOS requires root to bind port 53 — no USER directive
+ENTRYPOINT ["/coredns", "-conf", "/Corefile"]

mikrotik/coredns/plugin.cfg

@@ -0,0 +1,7 @@
+errors:errors
+log:log
+health:health
+template:template
+cache:cache
+forward:forward
+reload:reload