ci(test): host-persistent target dir; fast fmt gate; drop dead rust-cache

The Gitea actions cache on this instance stores caches (1.85 GB saves
confirmed in run 597) but never restores them — every run back through
run 584+ logs 'No cache found' even for exact keys saved an hour
earlier, including master-to-master. Every CI run has therefore been a
full cold build (~37 min), plus ~4 min tarring a cache nobody reads.

- Point CARGO_TARGET_DIR at a persistent path on the rust-host runner
  (host executor — filesystem carries over between runs), with a 40 GiB
  prune guard. Warm runs drop to minutes without touching the broken
  cache API.
- Drop Swatinem/rust-cache (pure overhead until the server is fixed).
- Split cargo fmt --check into a seconds-long fmt job gating the heavy
  test job, so a formatting slip can't burn a 35-minute build again
  (run 600).

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

.gitea/workflows/test.yml

@@ -2,6 +2,19 @@
 # locally, run on every master push and pull request. Until this workflow
 # existed, nothing in CI ran the test suite at all — a direct push to
 # master was entirely unguarded.
+#
+# Build caching (2026-07-13): the Gitea actions cache never restored on
+# this instance — every run back through run 597 logged "No cache found"
+# even for exact keys saved successfully ("Cache saved successfully") by
+# a run an hour earlier, including master→master restores. Until the
+# cache server on the runner host is fixed, Swatinem/rust-cache is pure
+# overhead here. `rust-host` is a HOST executor (its filesystem persists
+# between runs — ~/.cargo and the rustup toolchain already carry over),
+# so we get warm builds by pointing CARGO_TARGET_DIR at a persistent
+# path on the runner instead of tarring gigabytes through a cache API
+# that never returns them. Concurrent runs are safe: cargo serialises
+# on the target-dir lock.
+
 name: Test
 
 on:
@@ -28,8 +41,27 @@ on:
   workflow_dispatch:
 
 jobs:
+  # Seconds-long formatting gate in its own job so a rustfmt slip fails
+  # here instead of after a 35-minute cold build (run 600 spent its
+  # whole build budget to report an unformatted file).
+  fmt:
+    runs-on: rust-host
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install Rust 1.95.0
+        uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: 1.95.0
+          components: rustfmt
+
+      - name: Format check
+        run: cargo fmt --check
+
   test:
     runs-on: rust-host
+    needs: fmt
 
     # Full debuginfo made the solitaire_engine test-binary link peak past the
     # runner's memory — ld was OOM-killed (signal 9) on runs 447 and 486.
@@ -40,10 +72,19 @@ jobs:
     # test binaries concurrently; as the workspace grew (runs 514/516/519)
     # two+ simultaneous ld processes OOM-killed the runner again even at
     # line-tables-only. Two jobs keeps at most two links in flight — the
-    # compile-throughput cost is small next to the cache-warm build.
+    # compile-throughput cost is small next to the warm build.
+    #
+    # CARGO_INCREMENTAL=0: incremental artifacts bloat the persistent
+    # target dir for little benefit in CI (rust-cache used to set this
+    # for the same reason).
+    #
+    # CARGO_TARGET_DIR: persistent on the runner host — see the header
+    # comment. The prune step below keeps it from growing unbounded.
     env:
       CARGO_PROFILE_DEV_DEBUG: line-tables-only
       CARGO_BUILD_JOBS: '2'
+      CARGO_INCREMENTAL: '0'
+      CARGO_TARGET_DIR: /home/runner/.cache/ferrous-solitaire/target
 
     steps:
       - name: Checkout
@@ -53,10 +94,21 @@ jobs:
         uses: dtolnay/rust-toolchain@master
         with:
           toolchain: 1.95.0
-          components: clippy, rustfmt
+          components: clippy
 
-      - name: Cache cargo build
+      # Toolchain or lockfile bumps strand stale artifacts nothing will
-        uses: Swatinem/rust-cache@v2
+      # ever reuse; reset the dir when it crosses 40 GiB rather than
+      # curating it (a cold rebuild every few weeks is cheaper than the
+      # bookkeeping).
+      - name: Prune persistent target dir when oversized
+        run: |
+          limit_kb=$((40 * 1024 * 1024))
+          used_kb=$(du -sk "$CARGO_TARGET_DIR" 2>/dev/null | cut -f1 || echo 0)
+          echo "persistent target dir: $((used_kb / 1024)) MiB (limit $((limit_kb / 1024)) MiB)"
+          if [ "${used_kb:-0}" -gt "$limit_kb" ]; then
+            echo "over limit — clearing for a fresh cold build"
+            rm -rf "$CARGO_TARGET_DIR"
+          fi
 
       # Native link deps for the Bevy crates (engine/app/web) on a bare
       # ubuntu runner: ALSA + udev for input/audio, X11 + Wayland for winit.
@@ -67,9 +119,6 @@ jobs:
             libasound2-dev libudev-dev pkg-config libx11-dev libxcursor-dev \
             libxrandr-dev libxi-dev libwayland-dev libxkbcommon-dev
 
-      - name: Format check
-        run: cargo fmt --check
-
       # SQLX_OFFLINE uses the checked-in `.sqlx/` query cache (no live DB),
       # same as the web-e2e workflow's server prebuild.
       - name: Clippy (deny warnings)

solitaire_server/e2e/scripts/cycle_metrics.js

@@ -206,7 +206,9 @@ async function main() {
                     invariant_ok: !!snap?.invariants?.state_ok,
                     history_len: Array.isArray(snap?.move_history) ? snap.move_history.length : null,
                     replay_payload_present: payload !== null,
-                    replay_moves_len: Array.isArray(payload?.moves) ? payload.moves.length : 0,
+                    replay_moves_len: Array.isArray(payload?.recording?.instructions)
+                        ? payload.recording.instructions.length
+                        : 0,
                 };
             }, { stepCap: maxSteps, policyName: policy, maxVisits: maxVisitsPerState });
 

solitaire_server/e2e/tests/gameplay_review.spec.js

@@ -69,9 +69,9 @@ test("draw-mode toggle affects replay payload draw_mode", async ({ page }) => {
 
     const payload = await page.evaluate(() => window.__FERROUS_DEBUG__.replayPayload());
     expect(payload.draw_mode).toBe("DrawThree");
-    expect(payload.schema_version).toBe(2);
+    expect(payload.schema_version).toBe(4);
-    expect(Array.isArray(payload.moves)).toBeTruthy();
+    expect(Array.isArray(payload.recording?.instructions)).toBeTruthy();
-    expect(payload.moves.length).toBeGreaterThan(0);
+    expect(payload.recording.instructions.length).toBeGreaterThan(0);
 });
 
 test("autonomous play keeps invariants stable across seed batch", async ({ page }) => {

solitaire_server/e2e/tests/smoke.spec.js

@@ -47,7 +47,7 @@ test("debug failure report contains replay diagnostics", async ({ page }) => {
     expect(report.invariants).toBeTruthy();
 });
 
-test("replay payload builder exports schema-v2 moves", async ({ page }) => {
+test("replay payload builder exports a schema-v4 recording", async ({ page }) => {
     await page.goto("/play-classic?seed=42");
     await page.waitForFunction(() => typeof window.__FERROUS_DEBUG__ === "object");
 
@@ -57,10 +57,14 @@ test("replay payload builder exports schema-v2 moves", async ({ page }) => {
         .poll(async () => await page.evaluate(() => window.__FERROUS_DEBUG__.replayPayload() !== null))
         .toBe(true);
     const payload = await page.evaluate(() => window.__FERROUS_DEBUG__.replayPayload());
-    expect(payload.schema_version).toBe(2);
+    // Schema v4: the wasm layer assembles the whole payload; the deal is
+    // embedded in `recording` and moves live at recording.instructions.
+    expect(payload.schema_version).toBe(4);
     expect(payload.draw_mode).toMatch(/Draw(One|Three)/);
     expect(payload.mode).toBe("Classic");
-    expect(Array.isArray(payload.moves)).toBeTruthy();
+    expect(payload.recording).toBeTruthy();
-    expect(payload.moves.length).toBeGreaterThan(0);
+    expect(payload.recording.initial_state).toBeTruthy();
-    expect(payload.win_move_index).toBe(payload.moves.length - 1);
+    expect(Array.isArray(payload.recording.instructions)).toBeTruthy();
+    expect(payload.recording.instructions.length).toBeGreaterThan(0);
+    expect(payload.win_move_index).toBe(payload.recording.instructions.length - 1);
 });