ci(project): add Gitea Actions deploy workflow and docs

Deploy to the production host on push to master via a self-hosted act_runner
(host/shell executor): git fetch + reset --hard in /opt/tanabata, then
docker compose up -d --build. Shell-only steps, so the host needs just git and
docker — no node, no rsync.

docs/DEPLOY.md covers the one-time setup: what a runner is, the runner user,
cloning to /opt/tanabata with a read-only deploy key, registering act_runner
with the host label, and the host .env. Notes the security reason to scope the
runner to this repository.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.gitea/workflows/deploy.yml

@@ -0,0 +1,44 @@
+name: deploy
+
+# Build the image and (re)start the compose stack on the production host
+# whenever master moves. Also runnable manually from the Gitea Actions tab.
+on:
+  push:
+    branches: [master]
+  workflow_dispatch: {}
+
+# One deploy at a time; queue rather than cancel an in-flight run.
+concurrency:
+  group: deploy-prod
+  cancel-in-progress: false
+
+jobs:
+  deploy:
+    # Self-hosted act_runner registered on the prod host with the "host" label
+    # (shell executor), so the job uses the host's git + Docker daemon and the
+    # existing clone in /opt/tanabata. See docs/DEPLOY.md for runner setup.
+    #
+    # Only shell steps here (no `uses:` actions), so the host needs git + docker
+    # and nothing else — no node, no rsync.
+    runs-on: host
+
+    env:
+      DEPLOY_DIR: /opt/tanabata
+
+    steps:
+      - name: Pull latest master
+        # DEPLOY_DIR is a git clone set up once at deploy time. reset --hard
+        # makes it match origin exactly; .env is untracked (.gitignore) so it
+        # is never touched.
+        run: |
+          cd "$DEPLOY_DIR"
+          git fetch --prune origin
+          git reset --hard origin/master
+
+      - name: Build image and start the stack
+        working-directory: /opt/tanabata
+        # .env must already exist in DEPLOY_DIR on the host (secrets + DB mode).
+        run: docker compose up -d --build --remove-orphans
+
+      - name: Prune dangling build layers
+        run: docker image prune -f

docs/DEPLOY.md

@@ -0,0 +1,182 @@
+# Deployment (Gitea Actions → host)
+
+Tanabata is deployed by a [Gitea Actions](https://docs.gitea.com/usage/actions/overview)
+workflow ([`.gitea/workflows/deploy.yml`](../.gitea/workflows/deploy.yml)) that
+runs on the **production host itself**. On every push to `master` it updates the
+git clone in `/opt/tanabata` and runs `docker compose up -d --build` there, so the
+image is built from the freshly-pushed code and the stack is restarted.
+
+```
+push master ──> Gitea (container) ──> act_runner (host, "host" label)
+                                          │ git fetch + reset --hard   (in /opt/tanabata)
+                                          └ docker compose up -d --build
+```
+
+The Gitea server runs in a container, but the **runner runs directly on the host**
+(shell executor) so it can use the host's git, the host Docker daemon, and the
+clone in `/opt/tanabata`. Nothing needs a registry — the host builds the image
+locally. The workflow uses only shell steps, so the host needs just **git** and
+**docker** (no node, no rsync).
+
+## What is a runner?
+
+Gitea (like GitHub) only *coordinates* CI: it stores the workflow, queues jobs,
+and shows logs. It does **not** execute anything itself. A **runner** is a
+separate agent program that polls Gitea for queued jobs, runs the steps on a
+machine you control, and reports results back.
+
+Gitea's official runner is **act_runner** (a single Go binary; it uses the
+`act` engine to interpret workflow YAML). One act_runner process can serve many
+repos. Each runner advertises one or more **labels**, and a job's `runs-on:`
+picks a runner by label. A label also decides *how* a job runs — the **executor**:
+
+- **docker executor** — each job runs in a fresh container from an image (e.g.
+  `node:20-bookworm`). Isolated and reproducible; the usual default. Label form
+  at registration: `ubuntu:docker://node:20-bookworm`.
+- **host / shell executor** — the job runs directly on the host as the runner's
+  user, using host-installed tools. Label form: `host:host`. This is what we use,
+  because the deploy needs the host's Docker daemon and `/opt/tanabata`.
+
+So `runs-on: host` in the workflow ⇒ "run this job on a runner that registered a
+`host` label" ⇒ our shell executor on the prod box.
+
+## One-time setup
+
+### 1. Enable Actions in Gitea
+
+Gitea 1.21+ has Actions on by default. Otherwise add to `app.ini` and restart:
+
+```ini
+[actions]
+ENABLED = true
+```
+
+### 2. A runner user on the host
+
+Pick (or create) the Linux user the runner runs as. It must be able to use Docker
+and own the deploy dir — so the workflow needs no `sudo`:
+
+```bash
+sudo useradd -r -m -d /home/gitea-runner gitea-runner   # or reuse an existing user
+sudo usermod -aG docker gitea-runner                     # host Docker access
+```
+
+The host needs `git` and a Docker engine with the Compose plugin:
+
+```bash
+sudo apt install -y git docker.io docker-compose-plugin   # Debian/Ubuntu
+```
+
+### 3. Clone the repo to /opt/tanabata once
+
+The workflow only does `git fetch` + `reset --hard`, so the clone (and its auth)
+is established here, once. Use a **read-only deploy key** so the host never holds
+write credentials:
+
+```bash
+# As the runner user, create a key and add the PUBLIC half to the repo in Gitea:
+#   Repo → Settings → Deploy Keys → Add (read-only)
+sudo -u gitea-runner ssh-keygen -t ed25519 -f /home/gitea-runner/.ssh/tanabata_deploy -N ''
+
+# Clone with that key (SSH URL of your Gitea repo):
+sudo -u gitea-runner GIT_SSH_COMMAND='ssh -i /home/gitea-runner/.ssh/tanabata_deploy' \
+  git clone git@gitea.example.com:you/tanabata.git /opt/tanabata
+sudo chown -R gitea-runner:gitea-runner /opt/tanabata
+```
+
+> HTTPS works too — clone with a URL that carries a read-only token. SSH deploy
+> keys are the cleaner, per-repo, read-only option.
+
+After cloning, recurring `git fetch` reuses the remote + key stored in
+`/opt/tanabata/.git/config`, so the runner itself needs no standing credentials.
+
+### 4. Register and run act_runner on the host
+
+Get a registration token in Gitea. **Where you create it sets the runner's
+scope** (and `--name` is only a display label, unrelated to scope):
+
+- **Repository** (Tanabata repo → Settings → Actions → Runners) → serves only
+  this repo. **Use this.**
+- Organization → all repos in the org; Site (admin) → all repos on the instance.
+
+> Security: this runner is a host/shell executor with access to the Docker
+> socket — effectively root on the host. Register it at the **repository** level
+> so only Tanabata's workflows can run on your prod server; a site-wide runner
+> would let any repo's workflow execute arbitrary commands here.
+
+Then, as the runner user:
+
+```bash
+# Download act_runner: https://gitea.com/gitea/act_runner/releases
+act_runner register --no-interactive \
+  --instance https://gitea.example.com \
+  --token   <REGISTRATION_TOKEN> \
+  --name    prod-host \
+  --labels  host:host          # <-- maps `runs-on: host` to the shell executor
+
+# Run it (use a systemd unit in production so it survives reboots):
+act_runner daemon
+```
+
+`--labels host:host` is what makes jobs run **on the host** instead of in a
+container. The instance URL must be reachable from the host (Gitea's published
+port / domain — not the in-container address). Registration writes a `.runner`
+file (the runner's credentials) in the working directory.
+
+Minimal systemd unit (`/etc/systemd/system/act_runner.service`):
+
+```ini
+[Unit]
+Description=Gitea act_runner
+After=docker.service
+Requires=docker.service
+
+[Service]
+User=gitea-runner
+WorkingDirectory=/home/gitea-runner
+ExecStart=/usr/local/bin/act_runner daemon
+Restart=always
+
+[Install]
+WantedBy=multi-user.target
+```
+
+```bash
+sudo systemctl enable --now act_runner
+```
+
+### 5. Create /opt/tanabata/.env (secrets)
+
+The workflow **never** writes `.env` — it lives on the host and holds the real
+secrets and the chosen DB mode. `.env` is git-ignored, so `git reset --hard`
+leaves it untouched. Create it once:
+
+```bash
+cd /opt/tanabata
+sudo -u gitea-runner cp .env.example .env
+sudo -u gitea-runner $EDITOR .env    # set JWT_SECRET, ADMIN_PASSWORD, DATABASE_URL, etc.
+```
+
+See [`.env.example`](../.env.example) for every variable. For the bundled
+Postgres keep `COMPOSE_PROFILES=with-db`; to use a Postgres already on the host,
+set it empty and point `DATABASE_URL` at `host.docker.internal`.
+
+> Data lives in named Docker volumes by default (or the `*_DIR` host paths you
+> set in `.env`, e.g. `/var/lib/tanabata/...`) — **not** in `/opt/tanabata`. So
+> `git reset --hard` on the code dir never touches your data.
+
+## Deploying
+
+Push to `master` (or hit **Run workflow** on the Actions tab). Watch progress
+under the repo's **Actions** tab. The first build pulls the Node/Go base images
+and takes a few minutes; later builds reuse the host's layer cache.
+
+## Notes / alternatives
+
+- **Docker-executor runner instead of host.** If you'd rather the runner itself
+  run in a container, register with a Docker label and bind-mount
+  `/var/run/docker.sock` and `/opt/tanabata` into the job (act_runner
+  `config.yaml` → `container.valid_volumes`), then change `runs-on` accordingly.
+  The host executor above is simpler for host deploys.
+- **Zero-downtime** isn't attempted: `compose up` recreates changed containers.
+  For a single-node setup the brief restart is usually fine.