lb: trust the host Caddy's X-Forwarded-Proto

Every OAuth redirect URI the backend builds is currently http://, whatever
scheme the user actually arrived with.

There are two Caddy hops in front of the app: the host Caddy terminates TLS
and sets X-Forwarded-Proto: https, then proxies to 127.0.0.1:8137 over plain
HTTP; the stack LB is the second hop. Caddy preserves an incoming
X-Forwarded-* header only when the immediate peer is a trusted proxy, and
trusted_proxies was never set here -- so the LB overwrote the header with the
scheme of the connection it had just received, which is http.

accounts/oauth.py:_redirect_uri builds the callback from x-forwarded-proto,
and that URI must match what is registered in the provider's console.

Measured on vps2, same request to each hop:

  direct to web, XFP=https      -> https://thermograph.org/api/v2/...
  through the LB, XFP=https     -> http://thermograph.org/api/v2/...
  through the PATCHED LB        -> https://thermograph.org/api/v2/...
  through the PATCHED LB, no XFP-> http://thermograph.org/api/v2/...

The last line is the point of using trusted_proxies rather than hardcoding
`header_up X-Forwarded-Proto https`: the LB still reports the truth when
nothing in front of it claims otherwise.

private_ranges covers 127.0.0.1/8 and the Docker bridge ranges, which is where
the host Caddy reaches this container from. The LB binds loopback only,
precisely so it cannot be reached un-fronted, so the only party that can set
these headers is the host Caddy -- trusting the local peer widens nothing.

Applied to beta's LB too. Beta is where a provider change gets tested before
it reaches thermograph.org, so it needs the same behaviour or the test is not
a test.

.forgejo/workflows/infra-sync.yml

@@ -163,3 +163,105 @@ jobs:
               render_thermograph_secrets /opt/thermograph/infra prod /etc/thermograph.env
             fi
             echo "synced to $(git log --oneline -1)"
+
+  sync-centralis:
+    # Centralis' /etc/centralis.env has been HAND-MAINTAINED. The vault has held a
+    # copy since deploy/secrets/centralis.prod.yaml was seeded, and
+    # render_centralis_secrets has existed since the OpenBao work -- but nothing
+    # on any deploy path ever called it. A renderer with no caller is not a
+    # migration, it is a file.
+    #
+    # That gap is why enabling Google sign-in on Centralis had no safe home: the
+    # only way to set CENTRALIS_GOOGLE_CLIENT_ID was to hand-edit the live file,
+    # which is exactly what render-secrets.sh's header warns against.
+    #
+    # WHY THIS JOB RECREATES A CONTAINER WHEN THE OTHER THREE DELIBERATELY DO NOT.
+    # This workflow's header says it never rolls a service, because image tags are
+    # the app domains' axis. Centralis is different in kind: its compose file
+    # interpolates /etc/centralis.env at PARSE time (`set -a && . /etc/centralis.env
+    # && docker compose up -d`), so rendering the file without recreating the
+    # container changes nothing at all. Rendering and not recreating would be the
+    # silent no-op this estate keeps getting bitten by. It is one container with no
+    # replicas -- a two-second recreate, not a deploy.
+    if: github.ref_name == 'main'
+    runs-on: docker
+    concurrency:
+      group: infra-sync-centralis
+      cancel-in-progress: false
+    steps:
+      - name: Render /etc/centralis.env from the vault, then recreate Centralis
+        uses: https://github.com/appleboy/ssh-action@v1.2.0
+        with:
+          host: ${{ secrets.VPS2_SSH_HOST }}
+          username: ${{ secrets.VPS2_SSH_USER }}
+          key: ${{ secrets.VPS2_SSH_KEY }}
+          port: ${{ secrets.VPS2_SSH_PORT }}
+          script: |
+            set -euo pipefail
+            umask 077
+            cd /opt/thermograph
+            [ -f infra/deploy/secrets/centralis.prod.yaml ] || { echo "no Centralis vault — nothing to do"; exit 0; }
+
+            . infra/deploy/env-topology.sh
+            thermograph_topology prod
+            . infra/deploy/render-secrets.sh
+
+            scratch=$(mktemp -d); chmod 700 "$scratch"
+            trap 'rm -rf -- "$scratch"' EXIT
+
+            # Render to scratch FIRST. CENTRALIS_RENDER_DEST is the override the
+            # renderer already carries for exactly this: look before writing.
+            CENTRALIS_RENDER_DEST="$scratch/rendered.env" \
+              render_centralis_secrets /opt/thermograph/infra
+
+            # THE GUARD. On 2026-07-24 this estate lost Centralis' token registry by
+            # writing a file that was missing a key the live one had. The renderer
+            # proves its OWN output round-trips through bash, but it cannot know
+            # about a key that exists only in the live file -- someone hand-editing
+            # /etc/centralis.env to add, say, CENTRALIS_GOOGLE_CLIENT_ID is invisible
+            # to it. So: the vault must be a SUPERSET of what is live, or nothing is
+            # written and the job fails loudly.
+            #
+            # Names only. Both files are sourced in a clean environment so the
+            # comparison is over what bash actually ends up with, not over text.
+            # The `|| true` is load-bearing under `set -o pipefail`: grep exits 1 when
+            # a file yields no matching keys, which would otherwise abort the job on
+            # the exact edge case this guard exists to handle -- an empty or
+            # unreadable live file, i.e. a first run.
+            keys() { env -i bash -c 'set -a; . "$1" >/dev/null 2>&1; set +a; compgen -v' _ "$1" \
+                     | { grep -E '^(CENTRALIS_|DISCORD_)' || true; } | sort -u; }
+            keys "$scratch/rendered.env" > "$scratch/new.keys"
+            if [ -r /etc/centralis.env ]; then keys /etc/centralis.env > "$scratch/live.keys"; else : > "$scratch/live.keys"; fi
+
+            missing=$(comm -23 "$scratch/live.keys" "$scratch/new.keys" || true)
+            if [ -n "$missing" ]; then
+              echo "!! REFUSING to write /etc/centralis.env."
+              echo "!! The live file has keys the vault does not, so rendering would DELETE them:"
+              printf '!!   %s\n' $missing
+              echo "!! Someone hand-edited the live file. Put those keys in the vault instead:"
+              echo "!!   sops edit infra/deploy/secrets/centralis.prod.yaml"
+              exit 1
+            fi
+
+            added=$(comm -13 "$scratch/live.keys" "$scratch/new.keys" || true)
+            [ -n "$added" ] && printf '    new key from the vault: %s\n' $added
+
+            # Same file, byte for byte? Then there is nothing to do and no reason to
+            # bounce a healthy container.
+            if [ -r /etc/centralis.env ] && cmp -s "$scratch/rendered.env" /etc/centralis.env; then
+              echo "==> /etc/centralis.env already matches the vault — not recreating"
+              exit 0
+            fi
+
+            install -m 0600 -o "$(stat -c %U /etc/centralis.env 2>/dev/null || echo "$USER")" \
+                    -g "$(stat -c %G /etc/centralis.env 2>/dev/null || echo "$USER")" \
+                    "$scratch/rendered.env" /etc/centralis.env 2>/dev/null \
+              || sudo install -m 0600 -o agent -g agent "$scratch/rendered.env" /etc/centralis.env
+            echo "==> wrote /etc/centralis.env from the vault"
+
+            # Recreate so compose re-reads it. `docker restart` would NOT: the values
+            # are interpolated when the compose file is parsed, not read at runtime.
+            cd /opt/centralis/deploy
+            set -a; . /etc/centralis.env; set +a
+            docker compose up -d
+            docker compose ps --format '{{.Name}}\t{{.State}}\t{{.Status}}'

infra/openbao/verify-parity.sh

@@ -36,8 +36,13 @@ usage: verify-parity.sh (--all | --env <dev|beta|prod> ...)
 Renders each environment through BOTH backends and compares the resulting
 KEY=value sets. Prints key names and counts only, never values.
 
-Exit status: 0 = every requested environment is at parity; 1 = a mismatch.
-Suitable as a CI / cron gate.
+--all means "every environment that lives on THIS host", not all three: dev is
+on vps1, beta and prod are on vps2, so no single box can check them all. An
+environment with no checkout here is skipped out loud. Skipping every one of
+them is an error, not a pass.
+
+Exit status: 0 = every environment checked here is at parity; 1 = a mismatch,
+or nothing was checkable. Suitable as a CI / cron gate.
 EOF
   exit 2
 }
@@ -69,17 +74,43 @@ done
 . "$INFRA_DIR/deploy/render-secrets-openbao.sh"
 
 overall=0
+checked=0
 for env_name in "${TARGETS[@]}"; do
-  echo "== parity check: $env_name"
-
   # Per-environment topology: TG_BAO_APPROLE (beta's differs, because beta shares a
-  # filesystem with prod) and TG_SKIP_COMMON. Called inside the loop because --all
-  # walks three environments and the values differ per environment.
+  # filesystem with prod), TG_SKIP_COMMON and TG_APP_DIR. Called before the header
+  # line because --all walks three environments and may skip some of them.
   if ! thermograph_topology "$env_name"; then
     echo "!! unknown environment: $env_name" >&2
     overall=1; continue
   fi
 
+  # NO ENVIRONMENT CAN BE CHECKED FROM A HOST IT DOES NOT LIVE ON, and --all used to
+  # pretend otherwise. dev is on vps1; beta and prod are on vps2. Running --all
+  # anywhere therefore guaranteed a failure on whichever environments are elsewhere,
+  # and the failure was actively misleading: on vps1, prod resolves TG_BAO_APPROLE to
+  # /etc/thermograph/openbao-approle, which on THAT box holds dev's credential, so
+  # prod's check authenticated as tg-dev and took a 403. Fails closed, but reads like
+  # a policy bug rather than "wrong box".
+  #
+  # Identity comes from the ENVIRONMENT'S OWN ADDRESS being bound here, not from its
+  # checkout existing. The first version of this check tested for TG_APP_DIR and was
+  # wrong: vps1 still carries a stale /opt/thermograph from before the estate split,
+  # so prod looked resident there and was checked anyway. A leftover directory is
+  # exactly the kind of thing that outlives the arrangement it belonged to.
+  #
+  # Skipping is announced, never silent. A gate that quietly narrows what it checked
+  # is worse than one that fails, because the run still goes green.
+  if [ -n "${TG_SSH_HOST:-}" ] && command -v ip >/dev/null 2>&1; then
+    if ! ip -4 -o addr show 2>/dev/null | awk '{print $4}' | cut -d/ -f1 \
+         | grep -qxF "$TG_SSH_HOST"; then
+      echo "== parity check: $env_name — SKIPPED, lives on ${TG_HOST:-another host} (${TG_SSH_HOST}), not here"
+      continue
+    fi
+  fi
+
+  echo "== parity check: $env_name"
+  checked=$((checked + 1))
+
   # dev renders WITHOUT common on both backends. On the SOPS side that is a caller
   # flag; on the OpenBao side the policy also forbids it. Taking the flag from the
   # topology rather than re-deriving it here keeps the comparison apples-to-apples
@@ -167,11 +198,23 @@ for k in sorted(order):
   cleanup
 done
 
+# Zero environments checked is a FAILURE, not a pass. Otherwise a --all run on a
+# host that carries none of them — a new box, a renamed checkout, a typo'd --env —
+# exits 0 having verified nothing, and a green run is exactly what the 7-day gate
+# counts. "Nothing was wrong" and "nothing was examined" must not look alike.
+if [ "$checked" -eq 0 ]; then
+  echo
+  echo "!! nothing was checked: none of [${TARGETS[*]}] has a checkout on this host." >&2
+  echo "!! dev lives on vps1; beta and prod live on vps2. Run this where they are." >&2
+  exit 1
+fi
+
 if [ "$overall" = 0 ]; then
   echo
-  echo "PARITY OK for: ${TARGETS[*]}"
+  echo "PARITY OK — ${checked} environment(s) checked on this host."
   echo "Safe to consider flipping TG_SECRETS_BACKEND for these environments."
-  echo "Recommended gate before prod: 7 consecutive green runs on all three."
+  echo "Gate before prod: 7 consecutive green runs covering ALL THREE environments."
+  echo "That means both hosts — a green run here says nothing about the other box."
 else
   echo
   echo "PARITY FAILED — do NOT flip TG_SECRETS_BACKEND." >&2