tests: fix APP_URL port mismatch in the interactive VM

`just boot-vm` forwards the guest's port 80 to host port 8080
(QEMU_NET_OPTS in the justfile), but the backend builds absolute URLs
(e.g. the stored company logo path returned by the upload API) from
APP_URL, which defaults to plain "http://localhost" with no port. The
SPA's own relative URLs are fine - they inherit whatever port the
browser is already on - but any backend-generated absolute URL then
points at host port 80, where nothing is listening, instead of the
forwarded 8080. That gave a silent net::ERR_CONNECTION_REFUSED for the
logo image specifically, even though the upload itself succeeded.

Verified directly: with APP_URL overridden to "http://localhost:8080",
a freshly uploaded logo's URL is
"http://localhost:8080/storage/.../....png" instead of missing the
port. Automated test checks are unaffected (checks derivation hash
unchanged) since they curl the guest's own localhost directly, with no
host-side port remap involved.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

CLAUDE.md

@@ -39,7 +39,8 @@ This project uses `just` (see `justfile`) and Nix directly.
   and disables `boot.loader.grub` purely to satisfy NixOS's built-in assertions when `nix flake check` evaluates
   `config.system.build.toplevel` directly — the qemu-vm module overrides both with a tmpfs root and direct kernel
   boot when actually building `config.system.build.vm` (what `just build-vm` uses), so these values are otherwise
-  unused.
+  unused. Also raises `virtualisation.cores`/`memorySize` to 2/2560 (qemu-vm's defaults, 1 core/1024MB, are too
+  tight for real interactive use — see the module architecture notes below on PDF generation's memory cost).
 - `nixos-test.nix` — the automated test (`pkgs.testers.nixosTest`, exposed as `checks.<system>.invoiceninja` in
   `flake.nix`). Builds its own `pkgs` (with `allowUnfree` baked in — `nixosTest` forbids setting `nixpkgs.config`
   on a node once `pkgs` is externally supplied) and boots two nodes from `common.nix` — one with
@@ -48,9 +49,12 @@ This project uses `just` (see `justfile`) and Nix directly.
   against each node: waits for mysql/redis/phpfpm/queue-worker/proxy to be active, checks the
   `invoiceninja-data-setup` marker files (`.initial-migration`, `.db-seeded`, `.admin-created`) to confirm the
   one-time setup actually completed (not just that the oneshot unit exited), curls `/` for the app shell, exercises
-  the `invoiceninja-manage` wrapper, and logs in via `/api/v1/login` using the test admin credentials as an
+  the `invoiceninja-manage` wrapper, logs in via `/api/v1/login` using the test admin credentials as an
-  end-to-end check that migrations/seeding/account creation all worked. This is the only automated test in the
+  end-to-end check that migrations/seeding/account creation all worked, checks the full language list is
-  repo — there is no separate unit-test suite.
+  available (see the `cache:forget languages` note above), and exercises PDF generation end-to-end (POSTs to
+  `/api/v1/preview` with no entity, which falls back to a self-contained mock PDF built via the same
+  chromium/snappdf pipeline real invoices use, and checks for the `%PDF` magic bytes — no client/invoice setup
+  needed). This is the only automated test in the repo — there is no separate unit-test suite.
 
 ## Architecture
 
@@ -123,7 +127,11 @@ expect Nix to report the correct hash on a failed build (or use `nix-prefetch`/`
     `127.0.0.1:80`/`[::1]:80` too — otherwise visiting the VM via the IP literal instead of `localhost` matches no
     vhost and Caddy silently returns an empty `200 OK` (a blank page with no error anywhere).
   - Chromium is required at runtime for PDF generation (`PDF_GENERATOR=snappdf`, `SNAPPDF_CHROMIUM_PATH`) and is
-    added to the PHP-FPM and data-setup services' `path`.
+    added to the PHP-FPM and data-setup services' `path`. Each PDF request launches a full chromium instance (not
+    a lightweight renderer), and PHP-FPM's default `pm.max_children = 8` means up to 8 of these can run
+    concurrently — on a memory-constrained deployment this can exhaust RAM (confirmed: even a couple of
+    concurrent PDF requests OOM-killed a 1024MB VM). Size deployments accordingly, or lower
+    `phpfpm.settings."pm.max_children"`.
 
 ### `flake.nix`
 

nixos-module/invoiceninja.nix

@@ -486,7 +486,6 @@ in
 
       recommendedTlsSettings = (if (cfg.hostname == "localhost") then false else true);
       recommendedGzipSettings = true;
-      recommendedProxySettings = true;
       recommendedOptimisation = true;
 
       clientMaxBodySize = cfg.maxUploadSize;

tests/default.nix

@@ -4,6 +4,9 @@
 {
   imports = [
     (modulesPath + "/profiles/qemu-guest.nix")
+    # Not imported by default for a plain nixosSystem in this nixpkgs (only nixosTest nodes get it
+    # automatically) - needed explicitly so virtualisation.cores/memorySize below are real options.
+    (modulesPath + "/virtualisation/qemu-vm.nix")
     (import ./common.nix { inherit proxyServer; })
   ];
 
@@ -26,4 +29,27 @@
   };
 
   services.resolved.enable = true;
+
+  # `just boot-vm` forwards the guest's port 80 to host port 8080 (QEMU_NET_OPTS in the justfile),
+  # but the backend builds absolute URLs (e.g. the stored company logo path) from APP_URL, which
+  # defaults to plain "http://localhost" with no port. The SPA's own relative URLs are fine (they
+  # inherit whatever port the browser is already on), but any backend-generated absolute URL then
+  # points at host port 80 - where nothing is listening - instead of the forwarded 8080, giving a
+  # silent net::ERR_CONNECTION_REFUSED for things like the logo image. Override to match the actual
+  # forwarded port.
+  services.invoiceninja.settings.APP_URL = "http://localhost:8080";
+
+  # The qemu-vm module's defaults (1 core, 1024MB) are too tight for real interactive use: MySQL +
+  # Redis + PHP-FPM + chromium (PDF generation, spawned per request - each is a full multi-process
+  # browser, not a lightweight renderer) running concurrently can exhaust 1GB fast, and this VM
+  # boots with panic_on_oom - so it's not a graceful slowdown, the whole VM goes down. Confirmed by
+  # reproducing a "logo upload hangs" report this way: at the 1024MB default, even a couple of
+  # concurrent PDF-generation requests triggered `redis-server invoked oom-killer` and a kernel
+  # panic. 2 cores / 2560MB comfortably handles 2 concurrent PDF generations alongside normal
+  # traffic (verified: ~1.4GB still free afterward); a deliberate flood of many concurrent PDF
+  # requests can still exhaust even this, since each is a full chromium instance - if that's a
+  # concern for your deployment, consider capping `phpfpm.settings."pm.max_children"` below its
+  # default of 8, or provisioning more RAM.
+  virtualisation.cores = 2;
+  virtualisation.memorySize = 2560;
 }

tests/nixos-test.nix

@@ -80,6 +80,17 @@ pkgs.testers.nixosTest {
             language_count = len(json.loads(statics)["languages"])
             assert language_count > 40, f"expected ~45 languages, got {language_count}"
 
+        with subtest("PDF generation works (chromium/snappdf)"):
+            # /api/v1/preview with no entity falls back to a self-contained mock PDF (no invoice/client
+            # setup needed), exercising the same chromium/snappdf pipeline real invoice PDFs use.
+            machine.succeed(
+                "curl -sf -X POST http://localhost/api/v1/preview "
+                + f"-H 'X-API-TOKEN: {token}' -H 'Content-Type: application/json' "
+                + "-d '{}' -o /tmp/preview.pdf"
+            )
+            magic_bytes = machine.succeed("head -c 4 /tmp/preview.pdf")
+            assert magic_bytes == "%PDF", f"expected a PDF, got: {magic_bytes!r}"
+
 
     nginx.start()
     caddy.start()