petal/deploy/README.md

Deploying Petal

Two deployments exist right now:

	host	shape	status
parodia	petal.parodia.dev / 100.64.0.1	docker compose behind the host's Traefik	canonical since 2026-07-27 — she signs in here
millenia	192.168.1.212 / 100.64.0.2	petal.service, bare binary on :8088, Piper as user systemd units	frozen fallback: a copy of her writing as it stood at the move, still owned by the pre-auth local user

Her writing moved to the VPS when sign-in landed (Phase 16/17): it is the instance that actually authenticates, its data directory is LUKS-encrypted, and it is reachable from anywhere. millenia was left running and untouched as a fallback — but the two diverge the moment anything is written on either, so it should be retired rather than kept in step. Everything below is the VPS side; the millenia Piper notes are kept in the appendix because that instance still runs them.

---

1. The VPS stack

docker-compose.yml at the repo root brings up three containers:

• petal — the single Go binary with the frontend embedded. Publishes no host port; Traefik is the only way in.
• piper-en / piper-zh — read-aloud. Each Piper HTTP server loads exactly one voice, so English and Chinese are separate containers off one image, with the models cached in a shared volume. They sit on an internal network with no published ports, so only Petal can reach them. Adding pt-PT in Phase 21 is a fourth service, not a new image.

They run as containers rather than the host systemd units millenia uses because Piper was never actually installed on the VPS, and the reala account has no lingering session to keep user units alive across logout.

Prerequisites on the host

• Docker with the compose plugin, and the existing external traefik network
• A DNS A record for the hostname pointing at the VPS (petal.parodia.dev is already in place)

First deploy

ssh reala@100.64.0.1
git clone https://gitea.parodia.dev/drwily/petal.git ~/petal
cd ~/petal
cp deploy/petal.env.example .env

Then edit .env:

• PETAL_UID / PETAL_GID — id -u / id -g for this account. ./data is a bind mount, so the image's own petal user has no claim on it; a mismatch shows up as unable to open database file (14) and a restart loop.
• AUTHENTIK_URL / AUTHENTIK_CLIENT_ID / AUTHENTIK_CLIENT_SECRET / PETAL_ALLOWED_SUBS — sign-in, see §4. Without them Petal runs as the single local user and must not be exposed.
• LLM_MODEL / LLM_CHAT_MODEL — see §3.

mkdir -p data/backups
docker compose up -d --build
docker compose ps        # all three healthy

Updating

cd ~/petal && git pull && docker compose up -d --build

The frontend is embedded in the binary, so a rebuild is the whole deploy. The client polls /api/version (a hash of the built index.html) and offers a refresh when it changes.

---

2. What Traefik does

Labels follow the convention the other services on this box use: the external traefik network, the web-secure entrypoint, the default cert resolver and compression@file. Petal adds its own response-header middleware (frame-ancestors 'self', HSTS, nosniff, Referrer-Policy: same-origin).

There is no auth middleware at the edge: Petal does its own (§4). /api/health and /api/version sit outside Petal's own auth for the same reason they always did — a monitoring probe must not need a session, and neither carries user data.

---

3. The LLM link over headscale

The vLLM backend stays on millenia and is reached over headscale (100.64.0.2). This is the only cross-VPN dependency, and by the LLM-minimalism principle it never gates essential functionality — spell check, gloss, vocabulary garden, search, export and read-aloud all keep working with the link down, and the status bar shows the warm 🌙 小助手在休息 · Petal's helper is resting · 文字已保存.

LLM_TIMEOUT is raised from the local-network default of 30s to 90s: the voice and collocation passes send a whole document, the timeout is a hard deadline on the completion call, and a WAN+VPN round trip eats the margin.

How the link is exposed — a forwarder, not a rebind

vLLM stays bound to 127.0.0.1:8000. vllm-headscale-proxy.service (a socat unit, in this directory) adds a second listener on 100.64.0.2:8000 that forwards to it.

The plan originally said to rebind vLLM itself. That turned out to be the expensive option: vllm-chat.service is shared — Petal, Gogobee and Open WebUI all point at 127.0.0.1:8000, and Open WebUI stores its endpoint in its own database rather than in env — so moving the bind address would mean editing three consumers and reloading a 35B AWQ model, minutes of downtime for all of them. The forwarder adds a door instead of moving one: local callers are untouched, and the only new exposure is on the VPN interface.

It binds 100.64.0.2 specifically, never 0.0.0.0: the far end of this link is a public host, and the LAN has no business seeing an unauthenticated inference endpoint.

sudo install -m 0644 deploy/vllm-headscale-proxy.service /etc/systemd/system/
sudo systemctl daemon-reload && sudo systemctl enable --now vllm-headscale-proxy
ss -lntp | grep 8000          # expect BOTH 127.0.0.1:8000 and 100.64.0.2:8000

The model id (qwen3.6-35b) goes into LLM_MODEL / LLM_CHAT_MODEL on the VPS. Verified end to end: a grammar checkpoint from petal.parodia.dev returns real suggestions in ~3s over the VPN.

---

4. Sign-in (Authentik OIDC)

Petal is an OIDC client in its own right: it runs the login itself rather than trusting a header from the proxy. Nothing about the container has to be unreachable for that to be safe.

Login turns on only when AUTHENTIK_URL, AUTHENTIK_CLIENT_ID and AUTHENTIK_CLIENT_SECRET are all set. With any of them missing Petal falls back to the single hardcoded local user — which is what local development wants, and what every deployment did before this landed. A host serving the public must have them set.

Register Petal in Authentik

In the Authentik admin UI (Applications → Providers → Create → OAuth2/OpenID Provider):

Field	Value
Client type	Confidential
Redirect URI	https://petal.parodia.dev/auth/callback (strict)
Scopes	openid, profile, email
Signing key	any (Petal fetches the JWKS from discovery)

Then create an Application bound to that provider, and copy the client id, the client secret, and the provider's OpenID Configuration Issuer (it looks like https://auth.parodia.dev/application/o/petal/ — the issuer, not the .well-known URL; Petal appends that itself).

Put them in .env:

AUTHENTIK_URL=https://auth.parodia.dev/application/o/petal/
AUTHENTIK_CLIENT_ID=…
AUTHENTIK_CLIENT_SECRET=…
PETAL_ALLOWED_SUBS=her@example.com,me@example.com

PETAL_ALLOWED_SUBS is the guest list: comma-separated OIDC subject ids and/or email addresses. Authentik fronts several applications on this host, and being a valid user there does not mean being a user here. Leaving it empty lets in everyone Authentik authenticates. Emails are accepted alongside subject ids precisely so the list can be written before anyone has logged in — a subject is an opaque uuid that doesn't exist until first sign-in.

A valid login that isn't on the list gets a warm bilingual "this Petal isn't yours to write in" page, and no account is provisioned.

Checking it

curl -si https://petal.parodia.dev/api/docs | head -1     # 401 without a session
curl -si https://petal.parodia.dev/auth/login | grep -i location   # → Authentik
docker compose logs petal | grep '^.*auth:'               # issuer + redirect at boot

The startup log prints the redirect URI it will use; if Authentik rejects the login with a redirect-uri mismatch, compare that line against what's registered.

Two things bit this deployment, both worth checking first if a login dies early:

• The issuer's trailing slash is significant. Authentik's is …/application/o/petal/, OIDC requires the discovered issuer to match the configured one byte-for-byte, and normalising the slash away makes discovery fail with did not match the issuer URL returned by provider.
• A provider created through the API or ak shell has an empty grant_types, which authentik reads as "no grant type is permitted here" and answers with invalid_request / The request is otherwise malformed before the login page ever appears. The admin UI fills the list in for you; scripted creation must set it (authorization_code, refresh_token).

Discovery is lazy and retried, so an Authentik outage blocks new logins but leaves existing sessions working — those only need Petal's own database.

Sessions

Opaque token in a petal_session cookie (HttpOnly, SameSite=Lax, Secure on https); the sessions table stores only its SHA-256, so a database copy yields nothing usable. Thirty-day sliding expiry — every request pushes it out, throttled to one write an hour. /auth/logout deletes the row, not just the cookie. Expired rows are pruned at startup.

To sign someone out everywhere immediately:

docker compose exec petal sh -c \
  "sqlite3 /data/petal.db \"DELETE FROM sessions WHERE user_id = '<sub>'\""

The edge gate is gone

Until Phase 16 there was a Traefik basic-auth middleware in front of everything, because Petal authenticated nobody and a public hostname was a public API. It was removed when OIDC went live on 2026-07-27, together with the separate unauthenticated /api/health router that existed only to escape it: every /api route now answers 401 without a session, and the only thing an anonymous visitor gets is the app shell and a redirect to sign in.

If you ever run this stack without AUTHENTIK_* configured — Petal then falls back to the single local user — put the gate back before pointing DNS at it:

traefik.http.routers.petal.middlewares: compression@file,petal-headers,petal-auth
traefik.http.middlewares.petal-auth.basicauth.users: ${PETAL_BASIC_AUTH:?}

with htpasswd -nbB petal 'your-password' in .env as PETAL_BASIC_AUTH.

---

4a. Moving an account (scripts/migrate_local_user.py)

Petal ran as one hardcoded user (users.id = 'local') before sign-in existed. Moving that writing onto a real account is a deliberate, one-off operation:

docker compose stop petal
python3 scripts/migrate_local_user.py data/petal.db --to <oidc-sub>     # dry run
python3 scripts/migrate_local_user.py data/petal.db --to <oidc-sub> \
    --email her@example.com --name "Her Name" --apply
docker compose up -d petal

The subject id is knowable before she has ever logged in. With authentik's default hashed_user_id sub mode it is the user's uid:

docker exec authentik-server-1 ak shell -c \
  "from authentik.core.models import User; print(User.objects.get(username='claire').uid)"

so the data can move first and she signs in to find it already there.

Start Petal once on the incoming database before migrating. A database carried over from another instance may be a schema behind, and the app applies migrations at startup; the script moves rows and does not touch the schema.

The script is dry-run by default, takes its own VACUUM INTO backup, runs as one transaction with foreign keys off, and verifies the row counts before it commits. It refuses to run while anything else has the database open, and refuses to merge into an account that already owns writing.

local comes back, and that's expected. db.Open seeds that row on every startup, so it reappears the moment Petal restarts after a migration. It owns nothing — the writing is on the real account — and it is only ever resolved to by StaticResolver, which a deployment with AUTHENTIK_* set never uses. Check SELECT COUNT(*) FROM documents WHERE user_id = 'local' if you want to be sure a migration took; the presence of the row itself says nothing.

---

4b. The dictionary (dict.db)

Word lookups for the French, European Portuguese and Spanish pairs come from DreamDict's built database, which Petal opens read-only beside petal.db. Petal imports DreamDict's dictionary package directly — there is no DreamDict service to run and nothing to reach over the VPN, which matters because a hover gloss must answer in milliseconds.

dict.db is optional. With no file at DICT_PATH Petal logs

dictionary: no dict.db at /data/dict.db — English/Chinese only

and serves lookups from the datasets compiled into the binary. The Chinese pair is unaffected either way — it stays on ECDICT (see below) — and a non-Chinese writer still gets English definitions, synonyms and pronunciation, losing only the translation. A dictionary that failed to deploy costs the gloss, not the popover. A file that is present but was never imported is a different matter and is logged as an error.

Installing it

The database is built by DreamDict's own import CLI from ~6 GB of source data; it is not built on the VPS. Copy the built file into the data volume:

# on the machine holding a built dict.db (millenia: ~/dreamdict/data/dict.db)
scp ~/dreamdict/data/dict.db reala@100.64.0.1:/home/reala/petal/data/dict.db
# on parodia
chown "$(id -u):$(id -g)" /home/reala/petal/data/dict.db
docker compose restart petal   # the handle is opened once, at startup

Expect ~450 MB. It sits inside the LUKS volume with everything else (§6). The backups name petal.db explicitly rather than sweeping the data directory (§5), so dict.db stays out of them — which is the right outcome and worth keeping: it is rebuildable from public data and would otherwise dominate every nightly snapshot. Petal never writes to it.

Why Chinese doesn't use it

The zh pair stays on the embedded ECDICT gloss, deliberately. Measured on the deployed database, DreamDict reaches a Chinese gloss for 53% of the 2,000 commonest English words; ECDICT covers essentially all of them and is in daily use by a real writer. lexicon.Set.For is where that decision lives — one switch, changed the day a comparison on her actual lookups says otherwise.

For pt-PT and French the same measurement reads 62%, which is why they use DreamDict: there is no alternative source for them at all.

Rebuilding it

Rebuilt 2026-07-27 to add Spanish (the previous file predated DreamDict's Spanish support). The recipe, since it will be needed again:

# on millenia, from a clean checkout of dreamdict main
./scripts/download-dict-data.sh ~/dreamdict/data     # idempotent; skips what's there
go run ./cmd/dictimport --data ~/dreamdict/data --db ./dict.db --clean

~6 minutes on 32 cores; the data directory is ~7 GB and mostly already downloaded. Build to a new path, never over a file in use — then verify by hash on both ends before swapping.

Two things worth knowing before trusting a rebuild:

• The SUBTLEX-US download fails (the source moved behind a manual export). It does not matter: the loader falls back to SUBTLEX-US.txt, which is present, and English "frequency" is mostly SCOWL's commonness bucket anyway — 1000/800/600/…/50, refined by SUBTLEX for only ~1,600 words. That is why the word-difficulty chip reads difficulty, not frequency.
• Check the other languages' counts are unchanged before shipping. The 2026-07 rebuild came out byte-identical for en/fr/pt-PT/zh, which is what says it added a language rather than quietly shifting the rest.

Gloss coverage of the 2,000 commonest English words, after the rebuild: es 68.6%, fr 63.1%, pt-PT 62.1%, zh 53.2%. The startup line reports actual per-language row counts, so a database missing a language says so.

---

5. Backups

On the VPS — folded into parodia-backup

Petal rides the host's existing offsite job (/usr/local/bin/parodia-backup, parodia-backup.timer, nightly ~03:40): age-encrypted to S3, 14-day retention, dead-man snitch. The host holds only the age public recipient, so it writes backups it cannot itself decrypt.

push petal.db.age       sqlite_file_dump /home/reala/petal/data/petal.db

/home/reala/petal/.env is in the same job's secrets tarball — it carries the interim basic-auth hash and, from Phase 16, the OIDC client secret.

Why sqlite_file_dump and not the script's existing sqlite_dump: that helper uses Python's iterdump, which does not reproduce an FTS5 virtual table. It emits documents_fts as a raw sqlite_master row plus its shadow tables, and replaying the result dies with no such table: documents_fts — verified by round-tripping a real dump on 2026-07-27. Petal's cross-document search would have been silently missing after any restore. sqlite_file_dump runs VACUUM INTO instead: a genuine database file, virtual tables intact, WAL folded in, no write lock. Restore is a copy rather than a replay.

If apply.db ever gains a virtual table, it needs the same treatment.

Restore (VPS)

age -d -i <offline-identity> petal.db.age > /tmp/petal.db   # from S3
cd ~/petal
docker compose stop petal                       # stop writers first
mv data/petal.db data/petal.db.before-restore   # keep the current state
rm -f data/petal.db-wal data/petal.db-shm       # a stale WAL against a new file
cp /tmp/petal.db data/petal.db
docker compose start petal
docker compose logs petal --tail 5              # expect "database ready"

To sanity-check an archive before committing to it, have Petal open it in a scratch directory — a clean exit means it reads end to end:

mkdir -p /tmp/restore-check && cp /tmp/petal.db /tmp/restore-check/petal.db
docker run --rm -v /tmp/restore-check:/data --user "$(id -u):$(id -g)" \
  --entrypoint sh petal:local -c '/app/petal -backup /data/verify.db'

On millenia — petal-backup.timer

Until 2026-07-27 her actual writing had no scheduled backup at all; the newest snapshot was a month old. It now runs nightly at 03:20 (Persistent=true, because the box isn't on 24/7 and a missed window would otherwise be skipped silently):

sudo install -m 0644 deploy/petal-backup.service deploy/petal-backup.timer /etc/systemd/system/
sudo systemctl daemon-reload && sudo systemctl enable --now petal-backup.timer
sudo systemctl start petal-backup.service     # prove it before trusting it

deploy/backup-petal.sh snapshots via petal -backup, gzips, age-encrypts with the parodia public recipient, pushes to the VPS over headscale with a post-transfer size check, and prunes both ends. The private identity is offline, so neither millenia nor the VPS can decrypt what it is holding — verified.

A manual snapshot any time, no tooling required:

cd ~/petal && ./petal -backup ~/petal/backups/manual-$(date -u +%Y%m%dT%H%M%SZ).db

Restore is a copy — stop Petal, drop the file in as data/petal.db, remove any stale -wal/-shm, start.

---

6. Encryption at rest

VPS — /home/reala/petal/data is a LUKS volume

deploy/setup-encrypted-data.sh puts the data directory on LUKS2 over a sparse file at /var/lib/petal-crypt.img. That covers petal.db, uploaded images/, and the TTS cache — which is synthesized audio of her sentences and is easy to forget.

LUKS-on-a-file rather than gocryptfs because Petal is SQLite in WAL mode: WAL needs a shared-memory index (-shm) mapped consistently across processes, and FUSE has a long history of subtle mmap/locking differences. A block device with ext4 behaves exactly like a disk to SQLite, which is the only guarantee worth having under a database.

What it protects, honestly. The key lives at /etc/petal/dataset.key on the same host so the volume auto-unlocks at boot. That is a deliberate availability tradeoff:

protects against	a decommissioned or resold disk; reading the raw block device; casual browsing of a filesystem snapshot that excludes /etc
does not protect against	anyone holding the whole VM image — they get the keyfile with the ciphertext; or anything at all while the host is running and mounted

Real protection from a provider-side snapshot needs the key off-box (fetched over the VPN at boot). Considered, not chosen.

Two things this setup got wrong the first time, both caught by rehearsing a reboot rather than trusting a clean run — worth knowing if you rebuild it:

• Mounting over a directory hides its contents, it does not remove them. The first pass left the original plaintext petal.db and WAL sitting on the unencrypted root filesystem, invisible under the mount. The script now shreds the originals before mounting and refuses to continue if the mountpoint will not come up empty.
• systemd-cryptsetup was not installed, so /etc/crypttab was ignored entirely and the volume would never have unlocked at boot. The script now refuses to run without the generator present.

Check it any time:

sudo ./deploy/setup-encrypted-data.sh --status

The mount-liveness guard

The mountpoint directory exists whether or not the volume is mounted, so a boot where the unlock failed would start Petal against an empty unencrypted directory and quietly serve a blank database — the failure that looks like data loss. data/.volume-ok lives on the encrypted filesystem and is bind-mounted with create_host_path: false, turning that into a loud container start failure:

Error response from daemon: invalid mount config for type "bind":
bind source path does not exist: /home/reala/petal/data/.volume-ok

Verified by unmounting and attempting a start.

A true reboot has not been tested — the VPS also runs matrix, lemmy, akkoma, gitea and authentik, so rebooting it is your call. The boot path was rehearsed through local-fs.target, which pulls the mount, which pulls the unlock.

millenia is not encrypted at rest

LVM, no LUKS. Her canonical writing sits in plaintext on the home box. Backups leaving it are age-encrypted; the disk itself is not.

---

7. Supervision and monitoring

Petal on millenia ran for months as a bare ./petal with PPID 1 — no unit, no screen session — so a crash or reboot left it down until someone noticed. It is now petal.service:

sudo install -m 0644 deploy/petal.service /etc/systemd/system/
sudo systemctl daemon-reload && sudo systemctl enable --now petal.service

Verified by kill -9-ing it and watching systemd bring it back.

Piper's silent-failure mode is fixed. Both units now set StartLimitIntervalSec=300 / StartLimitBurst=5. With RestartSec=3 and systemd's default 10-second window, only ~3 restarts ever landed inside it, so the burst limit was never reached and a dead service looped 26,800+ times over a day without ever entering failed. A genuinely broken Piper now shows up in systemctl --user --failed.

Still to do — an external probe

Nothing yet watches millenia from outside. uptime-kuma already runs on the VPS and can reach millenia over headscale, so the missing piece is two monitors (they need the uptime-kuma UI, hence not scripted here):

• http://100.64.0.2:8088/api/health — Petal itself
• a POST to http://100.64.0.2:8088/api/tts — catches a dead Piper, which /api/health will not, because read-aloud degrades silently to browser speech

---

Appendix — Piper on millenia (user systemd units)

millenia still runs Piper as user services; these are the original notes.

scp deploy/piper.service deploy/setup-piper.sh 192.168.1.212:/tmp/
ssh 192.168.1.212 'cd /tmp && sudo ./setup-piper.sh'

setup-piper.sh creates ~/piper/venv, installs piper-tts[http], downloads en_US-amy-medium into ~/piper/voices, installs and enables piper.service (loopback :5005), and smoke-tests it. Idempotent. Check it with systemctl status piper / journalctl -u piper -f.

Chinese runs as a second instance (piper-zh.service, :5006, zh_CN-huayan-medium):

scp deploy/piper-zh.service 192.168.1.212:~/.config/systemd/user/
ssh 192.168.1.212 'export XDG_RUNTIME_DIR=/run/user/$(id -u)
  ~/piper/venv/bin/python -m piper.download_voices zh_CN-huayan-medium --data-dir ~/piper/voices
  systemctl --user daemon-reload && systemctl --user enable --now piper-zh.service'

Petal's env then carries TTS_ENDPOINT=http://127.0.0.1:5005, TTS_ENDPOINT_ZH=http://127.0.0.1:5006 and the matching voice ids. The handler maps language → instance from config, so another language is another instance plus an env pair, no code change.

Adding a language (Phase 21 made this literal). Petal discovers its Piper instances from the environment: English is the unsuffixed TTS_ENDPOINT/TTS_VOICE_EN, and every other language is a TTS_ENDPOINT_<LANG>/TTS_VOICE_<LANG> pair. <LANG> is the base tag — PT, not PT_PT, because an environment variable name cannot hold a hyphen and only one Portuguese model is loaded regardless. Both halves must be set: an endpoint with no voice is dropped, so a half-finished language reads to the browser as "no voice here, use Web Speech" instead of erroring on every tap. The startup line names what it actually resolved:

read-aloud enabled (voices: en=en_US-amy-medium, pt=pt_PT-tugão-medium, zh=zh_CN-huayan-medium)

Portuguese: pt_PT-tugão-medium is the only European voice Piper ships. The other five pt_* models in the catalogue are all Brazilian, so the voice has to be named explicitly for the same reason the Hunspell dictionary did (Phase 21): the obvious default is the wrong country. Check what exists before assuming:

docker exec petal-piper-en python -c "import urllib.request,json; \
d=json.load(urllib.request.urlopen('https://huggingface.co/rhasspy/piper-voices/resolve/main/voices.json')); \
print([k for k in d if k.startswith('pt')])"

Slow replay. POST /api/tts takes slow: true, which raises Piper's length_scale to about 4/3 (≈0.75× pace). It is a separate cache entry, not a playback-rate trick, so the slow clip is synthesized once and then instant.

Piper version note: piper-tts moved synthesis from POST / to POST /synthesize in 1.6.0, with an identical request body. TTS_PATH selects which — it defaults to /, and both the VPS compose and millenia's start.sh now set /synthesize. If read-aloud starts returning 502 after a Piper upgrade, that flag is the fix.

The venv is fragile across Python upgrades. On 2026-07-27 millenia's Piper was found dead with 26,800+ failed restarts, silently since the Jul 26 reboot — read-aloud had been falling back to browser Web Speech the whole time. Root cause: an OS upgrade moved /usr/bin/python3 from 3.13 to 3.14, and venv/bin/python3 is a symlink to the system interpreter, so the venv's lib/python3.13/site-packages became invisible — sys.path contained no site-packages at all. The failure surfaced as the misleading No module named piper.http_server even though http_server.py was sitting right there on disk.

Fix (what was done — recreating the venv, not repairing it):

systemctl --user stop piper.service piper-zh.service
mv ~/piper/venv ~/piper/venv.broken-py313
python3 -m venv ~/piper/venv
~/piper/venv/bin/pip install "piper-tts[http]"
~/piper/venv/bin/python -c 'import piper.http_server'   # must not raise
systemctl --user start piper.service piper-zh.service

That reinstall lands 1.6.0, so it must be paired with TTS_PATH=/synthesize in start.sh and a binary new enough to read that variable. Voices in ~/piper/voices survive and do not need re-downloading.

Worth knowing: Restart=on-failure will retry forever without ever alerting. Neither service reports its health anywhere, which is why this went unnoticed for a day. A /api/tts probe in uptime-kuma would have caught it.

Verify end to end (through Petal, including the ffmpeg transcode):

curl -sf -X POST localhost:8088/api/tts \
  -H 'Content-Type: application/json' \
  -d '{"text":"hello there","lang":"en-US"}' -o /tmp/petal-tts.mp3 \
  && file /tmp/petal-tts.mp3   # expect: MPEG ADTS, layer III

A second identical call is served from the cache; a language with no configured instance returns 404 so the client falls back to Web Speech.