upgrade base: prefer the newest published release over a STALE canonical

The canonical is last-GREEN, not CURRENT. Its promotion (WC5) deploys a
warm-<recipe> app and health-checks it; when that fails the canonical stays
pinned at an old release indefinitely — and the failure is invisible to the
upgrade tier, which keeps basing on it.

Live evidence (2026-08-10): gitea's canonical was 3.5.3+1.24.2-rootless from
2026-06-17 because warm-gitea crash-loops on a read-only app.ini (open fix:
recipe PR #4). Four other recipes are in the same state — the 2026-08-09 sweep
logged GREEN-BUT-PROMOTE-FAILED for bluesky-pds, drone, gitea, lasuite-drive and
lasuite-meet.

Consequence: the upgrade tier tested 3.5.3 -> head, a transition no deployment
performs, and MISSED the one they do. 3.5.3 -> head crosses an APP_INI_VERSION
change (v21 -> v22) so Swarm creates a fresh config object; the real
3.6.1 -> 3.6.2 upgrade keeps v22 and aborts on Swarm's immutable-config rule
('only updates to Labels are allowed'). That shipped in 3.6.2 and broke every
existing install taking the 1.27.1 security release.

Now: when the newest published release older than head is NEWER than the
canonical, use the release and say so in the log (naming the stale canonical so
the promote failure gets noticed). Resolution stays dynamic; UPGRADE_BASE_FLOOR
still filters structurally-invalid candidates; the canonical remains the base
whenever it is current.

Unit-verified on the real gitea tag set: base moves 3.5.3+1.24.2-rootless ->
3.6.1+1.26.2-rootless, exactly the transition that reproduces the break.

runner/run_recipe_ci.py

@@ -175,6 +175,36 @@ def resolve_upgrade_base(
             flush=True,
         )
         rec = None
+    # STALE-CANONICAL GUARD (phase relbase): the canonical is only "last-green", NOT "current". Its
+    # promotion can fail for reasons unrelated to any PR (the WC5 promote deploys a warm-<recipe> app
+    # and health-checks it), and a failed promote leaves the canonical pinned at an OLD release
+    # indefinitely — gitea sat at 3.5.3+1.24.2 from 2026-06-17 to 2026-08-10 because warm-gitea
+    # crash-looped on a read-only app.ini. Basing the upgrade tier on that stale release tests a
+    # transition no deployment performs, and can silently MISS breaks in the transition users do
+    # perform: gitea 3.5.3→head crosses an APP_INI_VERSION change (v21→v22) so Swarm creates a fresh
+    # config, while the real 3.6.1→3.6.2 upgrade keeps v22 and aborts on Swarm's immutable-config
+    # rule. Prefer the newest published release older than head whenever it is newer than the
+    # canonical: that is what real installs upgrade from.
+    if rec and rec.get("version") and not skip_canonicals and head_version:
+        _rel_tags = warm_reconcile.recipe_tags(recipe)
+        if floor:
+            _rel_tags = [t for t in _rel_tags if not _below_floor(t)]
+        newest_rel = warm_reconcile.newest_older_version(_rel_tags, head_version)
+        if newest_rel and warm_reconcile.version_key(newest_rel) > warm_reconcile.version_key(
+            rec["version"]
+        ):
+            print(
+                f"== upgrade base: newest published release {newest_rel} is NEWER than the "
+                f"last-green canonical {rec['version']} — using the release (what deployments "
+                f"actually upgrade from); canonical is stale, check its WC5 promote",
+                flush=True,
+            )
+            return BasePlan(
+                "version",
+                newest_rel,
+                None,
+                f"newest published release older than head (canonical {rec['version']} is stale)",
+            )
     if rec and rec.get("version") and not skip_canonicals:
         canon = rec["version"]
         same = head_version is not None and warm_reconcile.version_key(

tests/STYLE.md

@@ -1,116 +0,0 @@
-# cc-ci test style guide
-
-Rules for writing and changing tests under `tests/`. Read this before any test edit — in particular
-before a `/recipe-upgrade <recipe> --with-tests` or `/ci-test-review` fix, where the temptation is to
-make a red run green rather than to make the test right.
-
-The tests are the **independent gate** on recipe upgrades. Their value is entirely in being hard to
-fool, so every rule below exists to keep them (a) honest and (b) alive across upgrades.
-
----
-
-## 1. Set up state through the application, not its database
-
-**Order of preference for any fixture that must create state:**
-
-1. **The app's public HTTP API.**
-2. **The app's official CLI or release console** (`docker exec … <app-cli>`).
-3. **Writing rows into its database — last resort only**, and only with a comment saying which of the
-   above were tried and why they did not work.
-
-Direct SQL couples the test to the app's *internal schema*, which upgrades are free to change. The
-app's own interface is the thing it promises to keep working.
-
-> **Why this rule exists.** `tests/plausible/custom/test_event_tracking.py` used to register its test
-> site with `INSERT INTO sites (...)`. That was sufficient for plausible v2. In v3 a site must belong
-> to a **team**, and the app silently discards events for a teamless site — `POST /api/event` still
-> returns **202** and the row is still in postgres, so the only visible symptom was that nothing ever
-> reached ClickHouse. It read as a mysterious ingestion stall and held the recipe RED for six weeks.
->
-> The fix was not to also INSERT a team row. It was to stop writing rows: the fixture now calls
-> `Plausible.Sites.create/2` through the app's release console, and the app provisions whatever its
-> data model currently needs. The same expression works unchanged on v2 (which has no `teams` table
-> at all) **and** v3 — not because the test handles both, but because it stopped depending on the
-> schema.
-
-When the ideal interface is unavailable, say so in the code. plausible's HTTP provisioning API
-(`POST /api/v1/sites`) is gated behind a paid plan and answers `:upgrade_required` on CE, so the test
-drops to option 2 and records that in a comment.
-
-## 2. Gate on version rather than writing dual-path fixtures
-
-If a behaviour genuinely only exists from version X, **gate the test on the version** instead of
-branching inside it:
-
-```python
-pytest.mark.skipif(app_version < (3,), reason="teams were introduced in v3")
-```
-
-Do **not** write a fixture that carefully supports both schemas. Version-portable code is harder to
-read, harder to trust, and quietly rots once nobody runs the old path.
-
-Corollary: **old tests can simply be deleted** once the fleet has moved past that version. The older
-version is only ever exercised through the *upgrade* tier (deploy base → upgrade → assert), so tests
-that only make sense for a superseded version are dead weight, not coverage.
-
-Prefer §1 first: an app-level fixture often makes the version difference disappear, and then no gate
-is needed at all.
-
-## 3. Never weaken an assertion to turn a run green
-
-There is a hard line between these two, and only the second is allowed as a way out of a red run:
-
-* **Weakening** — relaxing *what* is asserted: dropping a field check, accepting a wider status set,
-  asserting a 202 ack instead of the stored result, deleting the read-back.
-* **Correcting the fixture or the wait** — fixing *how* the test sets up or how long it allows, with
-  the assertion untouched.
-
-If a test can only pass by asserting less, it has found a real regression. Report it; do not edit it.
-
-## 4. Assert real state, not acknowledgements
-
-An HTTP 202 means "accepted", not "done". Read the effect back out of the system that owns it — the
-row in the analytics store, the file on disk, the record in the API — and assert on the values you
-sent. plausible's ingestion returns 202 for events it goes on to discard entirely; a test that
-stopped at the ack would have been permanently, silently green.
-
-## 5. Derive waits from the recipe's declared readiness, not a guess
-
-A per-recipe `recipe_meta.py` already declares `DEPLOY_TIMEOUT` / `HTTP_TIMEOUT` because someone
-measured that app's boot profile. A custom test that hard-codes a shorter window contradicts it and
-will flake or fail on a slower version.
-
-Remember the **tier order**: `custom` runs after `backup`/`restore`, which disrupts the datastore and
-restarts the app. A window sized for a warm app is not sized for that. plausible's health check
-allowed 60s; v3 boots through `sleep 10` → `createdb` → `migrate` → cache warmers first.
-
-## 6. Diagnose from the app's own telemetry before touching a test
-
-Before concluding a test is stale, find the app's account of what happened. It is usually definitive
-and it stops you fixing the wrong thing. plausible records dropped events in ClickHouse's
-`ingest_counters`: `dropped_not_found` with 0 rows before the fix, `buffered` with rows after — that
-single counter identified the root cause after the HTTP status had suggested everything was fine.
-
-Prove the diagnosis both ways where you can: same input, broken state → symptom; corrected state →
-no symptom.
-
-## 7. Fixtures must be idempotent
-
-A fixture may run against a warm canonical, a restored volume, or a re-run. Creating state must be
-safe to repeat — look the object up first and reuse it, rather than assuming a clean database.
-
-## 8. Keep test identities obviously synthetic
-
-Use `ccci-`-prefixed names and `.example` / `.invalid` domains for anything a test creates, so state
-it leaves behind is instantly attributable and can never be confused with real data.
-
----
-
-## Changing a test: the checklist
-
-1. Reproduce the failure and get the **app's own** explanation (§6).
-2. Classify: recipe bug, or stale test? Only a stale test justifies a test edit.
-3. Fix the **fixture, wait, or setup** — never the assertion (§3).
-4. Prefer the app's interface over its database (§1); gate on version rather than branching (§2).
-5. Verify green against the recipe PR head with the changed test, plus a regression sample.
-6. Say in the commit and PR **what evidence** proves the diagnosis, not just what changed.

tests/plausible/custom/test_event_tracking.py

@@ -14,11 +14,7 @@ Both assert real app state (the event reached the analytics store), not just the
 
 plausible only ingests events for *known* sites — the in-memory `sites_cache` gates ingestion and
 drops events for unregistered domains (empirically confirmed: an event for an unregistered domain
-never appears in events_v2). Sites are therefore provisioned through plausible's OWN creation path
+never appears in events_v2). So each test first registers a site row in the metadata postgres, then
-rather than by writing rows — under v3 a site must belong to a TEAM, and a teamless site is dropped as
-`dropped_not_found` while the POST still acks 202, which reads as a silent ingestion stall. Letting the
-app create the site sidesteps that entirely, and works unchanged on v2. So each test first provisions
-the site, then
 POSTs repeatedly while polling ClickHouse: the sites_cache must refresh to admit the new site and the
 event write-buffer must flush to ClickHouse, so the first landing is not instantaneous. Re-POSTing the
 same event is safe — we assert the row count is >= 1.
@@ -44,10 +40,6 @@ _UA = (
     "(KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36"
 )
 
-# Identity the harness provisions sites under. Ephemeral per-run deploy, never a real account.
-_HARNESS_EMAIL = "cc-ci@ci.invalid"
-_HARNESS_PW = "ccci-harness-passphrase-2026"
-
 
 def _ch(domain: str, sql: str) -> str:
     """Run a ClickHouse query against the `plausible_events_db` service; return stdout (stripped)."""
@@ -58,61 +50,18 @@ def _ch(domain: str, sql: str) -> str:
     ).strip()
 
 
-# plausible's own provisioning path. Creating a site through the app (rather than INSERTing rows)
-# means the app applies whatever its current data model requires — which is what makes this work
-# unchanged across the v2→v3 jump, where sites gained a mandatory owning TEAM. Verified on cc-ci
-# against BOTH v2.0.0 (no `teams` table at all) and v3.2.1: identical expression, site usable, events
-# ingested. See tests/STYLE.md.
-#
-# The HTTP provisioning API (`POST /api/v1/sites`) would be the first choice, but it is gated behind
-# a paid plan — on CE it answers `:upgrade_required` — so the app's release console is the closest
-# public interface available here.
-_PROVISION_SITE_EXS = """
-pw = "__PW__"
-email = "__EMAIL__"
-user =
-  case Plausible.Auth.find_user_by(email: email) do
-    nil ->
-      {:ok, u} =
-        Plausible.Auth.User.new(%{name: "cc-ci", email: email, password: pw, password_confirmation: pw})
-        |> Plausible.Repo.insert()
-      u
-    u -> u
-  end
-site = "__SITE__"
-result =
-  case Plausible.Sites.get_by_domain(site) do
-    nil -> Plausible.Sites.create(user, %{"domain" => site, "timezone" => "UTC"})
-    s -> {:ok, s}
-  end
-case result do
-  {:ok, _} -> IO.puts("CCCI_SITE_OK " <> site)
-  other -> IO.puts("CCCI_SITE_ERR " <> inspect(other))
-end
-"""
-
-
 def _register_site(domain: str, site: str) -> None:
-    """Provision `site` via plausible's own site-creation path, so it is a site the app will ingest for.
+    """Insert a site row into the metadata postgres (`db` service) so plausible will ingest events for
-
+    it. Idempotent (ON CONFLICT DO NOTHING)."""
-    Idempotent: an existing domain is reused rather than re-created.
+    sql = (
-
+        "INSERT INTO sites (domain, timezone, inserted_at, updated_at, native_stats_start_at) "
-    Do NOT reach into postgres to do this. A `sites` INSERT was enough under v2, but v3 requires the
+        f"VALUES ('{site}','UTC', now(), now(), now()) ON CONFLICT (domain) DO NOTHING; "
-    site to belong to a TEAM and silently discards events for a teamless site — `POST /api/event`
+        f"SELECT domain FROM sites WHERE domain = '{site}';"
-    still acks 202 and the row still exists, so the only symptom is that nothing reaches ClickHouse
-    (ClickHouse's own `ingest_counters` records it as `dropped_not_found`). That is what put this
-    recipe RED on build 1224. Going through the app removes the whole class of problem: it provisions
-    the team itself.
-    """
-    exs = (
-        _PROVISION_SITE_EXS.replace("__PW__", _HARNESS_PW)
-        .replace("__EMAIL__", _HARNESS_EMAIL)
-        .replace("__SITE__", site)
-    )
-    out = lifecycle.exec_in_app(domain, ["/app/bin/plausible", "rpc", exs], service="app")
-    assert f"CCCI_SITE_OK {site}" in out, (
-        f"could not provision site {site!r} via the app: {out.strip()[-400:]}"
     )
+    out = lifecycle.exec_in_app(
+        domain, ["psql", "-q", "-U", "plausible", "-d", "plausible", "-tAc", sql], service="db"
+    ).strip()
+    assert out == site, f"site {site!r} not registered in postgres (got {out!r})"
 
 
 def _post_event(base_domain: str, site: str, name: str, pathname: str) -> int:
@@ -144,9 +93,9 @@ def _ingest_and_count(
     last_status = None
     while True:
         last_status = _post_event(base_domain, site, name, pathname)
-        assert last_status == 202, (
+        assert (
-            f"POST /api/event for {name!r} → HTTP {last_status} (expected 202)"
+            last_status == 202
-        )
+        ), f"POST /api/event for {name!r} → HTTP {last_status} (expected 202)"
         time.sleep(interval)
         raw = _ch(base_domain, count_sql)
         count = int(raw) if raw.isdigit() else 0
@@ -194,6 +143,6 @@ def test_custom_event_roundtrip(live_app):
         live_app,
         f"SELECT name FROM events_v2 WHERE pathname = '{pathname}' LIMIT 1",
     )
-    assert stored_name == event_name, (
+    assert (
-        f"custom event stored as {stored_name!r}, expected {event_name!r}"
+        stored_name == event_name
-    )
+    ), f"custom event stored as {stored_name!r}, expected {event_name!r}"

tests/plausible/custom/test_health_check.py

@@ -17,12 +17,6 @@ def test_plausible_root_serves(live_app):
     62-char SECRET_KEY_BASE, see recipe_meta.EXTRA_ENV); the dedicated
     /api/health endpoint is.
     """
-    # The custom tier runs AFTER the backup/restore tier, which disrupts postgres under the app and
-    # restarts it. v3 (community-edition) then boots through `sleep 10` + `db createdb` + `db migrate`
-    # + cache warmers before /api/health flips to 200, which does not fit in 60s — that is what put
-    # this recipe RED on build 1224 while install/upgrade/backup/restore all passed. The assertion is
-    # unchanged (still a hard 200 from the real readiness endpoint); only the wait matches the boot
-    # profile the recipe already declares via recipe_meta.HTTP_TIMEOUT (1200).
     url = f"https://{live_app}/api/health"
-    status, _ = harness_http.retry_http_get(url, expect_status=(200,), max_wait=300, interval=5)
+    status, _ = harness_http.retry_http_get(url, expect_status=(200,), max_wait=60, interval=3)
     assert status == 200, f"GET {url} HTTP {status}"