One connection-settings store for every application

.github/workflows/ci.yml

@@ -8,10 +8,16 @@ on:
 
 jobs:
   build:
-    # qa-inspect-host is a native renderer binary. Its currently published
-    # Linux runtime predates Tauri's GTK trait additions, while macOS is the
-    # supported build used by the host's own CI and by AgencyZero.
-    runs-on: macos-14
+    # Linux, since `qa-inspect-host` 0.1.9. It could not be built here before:
+    # it reached the engine through `tauri-runtime-blitz`, which brought the
+    # Tauri runtime with it, and that does not compile on Linux at all -- its
+    # `RawWindow` is a `gtk::ApplicationWindow` there and a Blitz window has
+    # none. So this job ran on macOS to get a host at all.
+    #
+    # The host opens no window, and now does not compile one: it takes the
+    # runtime with `default-features = false`, so nothing in its graph is Tauri
+    # or GTK. macOS is for building packages, not for a component sweep.
+    runs-on: ubicloud-standard-2
     steps:
       - uses: actions/checkout@v4
 
@@ -50,9 +56,18 @@ jobs:
       # Check, Build and the whole test suite were skipped rather than failed —
       # a green-looking pipeline that had verified nothing.
       #
-      # `--locked` uses each crate's own published lockfile, which is the set
-      # its author tested. It costs nothing in currency: both tools still
-      # install at their newest version.
+      # `--locked` is gone, and it did cost currency: a lockfile is frozen at
+      # publication, so the host was pinned to whatever engine existed the day
+      # it was released. A renderer fix could then never reach this gate without
+      # a host release, and that is not hypothetical -- `ps-blitz` 0.4.3 fixes
+      # three defects in checkbox activation, and the host published before it
+      # kept installing 0.4.2, so `switch`, `checkbox` and `radio` went on
+      # failing here against an engine that had already been fixed.
+      #
+      # The mismatch `--locked` was added for is gone with it: it happened when
+      # `qa-inspect-host` wanted `ps-blitz` 0.4.x while its `tauri-runtime-blitz`
+      # wanted 0.3.x, so the same type existed twice. Both now require `^0.4`
+      # and resolve to one copy.
       #
       # One command per tool, deliberately. `cargo install a b --locked`
       # resolves the two together and downgrades to satisfy both — it picks
@@ -64,8 +79,16 @@ jobs:
           # which leaves rust-cache nothing useful to restore on the next run.
           CARGO_TARGET_DIR: target/qa-tools
         run: |
-          cargo install ps-qa --locked
-          cargo install qa-inspect-host --locked
+          # Pinned to a floor, not left to "latest".
+          #
+          # The sweep needs `--headless` and a host that can replace a text
+          # field without a font catalogue. Installing whatever is newest meant
+          # a version without those failed as `unexpected argument`, which
+          # reads as a broken workflow rather than a tool that is too old --
+          # and a host that silently appended to a field would not have failed
+          # at all until a check disagreed about a value.
+          cargo install ps-qa --version "^0.6.2"
+          cargo install qa-inspect-host --version "^0.1.12"
 
       - name: Install Dependencies
         run: bun install
@@ -91,16 +114,25 @@ jobs:
       # minutes while preserving a clean document between components.
       - name: Rendered component outcomes
         env:
-          # GitHub's shared macOS runner took 1.6-1.9s for outcomes that pass
-          # locally under 1.25s. This widens scheduling deadlines only; every
+          # A shared runner takes longer than a local machine for outcomes that pass
+          # locally well inside the budget. This widens scheduling deadlines only; every
           # declared rendered-state transition is still required.
           QA_TIMEOUT_SCALE: 2
+          # This runner has no font catalogue, and it should not have one: a
+          # component library that needs a GUI stack installed to be tested is
+          # a component library nobody can test. With no fonts every glyph
+          # shapes to nothing, so anything sized by its text lays out flat and
+          # every paint assertion fails for a reason that says nothing about
+          # the component. The headless profile asks the same questions about
+          # layout instead; the full profile keeps the visual half and runs
+          # where there are fonts. See `PROFILES` in generate-checks.ts.
+          QA_PROFILE: headless
         run: |
           bun run qa:checks
           bun run qa:entries
-          git diff --exit-code -- tests/ps-qa tests/qa-harness/entries
+          git diff --exit-code -- tests/ps-qa tests/ps-qa-headless tests/qa-harness/entries
           bun run qa:build
-          zsh tests/qa-harness/run-all.sh
+          bash tests/qa-harness/run-all.sh
 
       # Fails when docs/api-contract.md and the shipped types disagree, in
       # either direction. Never runs with --write: regenerating the document on