ci(test): stabilize v2 tests; move -race to a separate non-blocking job

The v2 test matrix has been red on newer Go versions and non-Linux
platforms. The suite has several pre-existing flaky data races (e.g. the
package-global fiber.Get client in Test_Client_UserAgent) and order-
dependent tests. Two things made this unfixable by CI config alone:

- A -race detector abort (os.Exit) fails the whole run, and neither the old
  nick-fields/retry (retries the whole run) nor gotestsum's per-test
  rerun-fails can mask it.
- gotestsum's rerun-fails is additionally defeated here because the recover
  test prints a literal "panic:" (via debug.Stack), which gotestsum treats
  as a suspected panic and then refuses to rerun any failed test.

So keep the whole-suite retry for the blocking job and take -race out of the
gate instead:

- Blocking Build job runs `go test ./... -count=1` (no -race, no -shuffle;
  -shuffle exposes the order-dependent tests) under nick-fields/retry, which
  has no panic heuristic to trip over.
- Add a separate non-blocking Race job (continue-on-error) that runs
  `go test -race -shuffle=on` so the pre-existing races stay visible without
  blocking merges. Fold -race back into Build once the races are fixed
  (#4496 fixes the first one).
- Raise the matrix floor to Go 1.20 to match the module's go.mod (go 1.20).
- Exclude Go <= 1.21 on macos-latest: those binaries abort on the arm64
  runner's newer dyld with "missing LC_UUID load command" (Go 1.22+ emits
  LC_UUID). Old Go stays covered on ubuntu and windows.
- Add fail-fast: false and explicit job timeouts.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: ReneWerner87

.github/workflows/test.yml

@@ -16,10 +16,19 @@ name: Test
 jobs:
   Build:
     strategy:
+      fail-fast: false
       matrix:
-        go-version: [1.17.x, 1.18.x, 1.19.x, 1.20.x, 1.21.x, 1.22.x, 1.23.x]
+        # Floor is 1.20 to match the module's go.mod (go 1.20).
+        go-version: [1.20.x, 1.21.x, 1.22.x, 1.23.x]
         platform: [ubuntu-latest, windows-latest, macos-latest]
+        exclude:
+          # Go <= 1.21 test binaries abort on macos-latest (arm64) with
+          # "dyld: missing LC_UUID load command"; Go 1.22+ emits LC_UUID.
+          # Old Go stays covered on ubuntu and windows.
+          - { go-version: 1.20.x, platform: macos-latest }
+          - { go-version: 1.21.x, platform: macos-latest }
     runs-on: ${{ matrix.platform }}
+    timeout-minutes: 20
     steps:
       - name: Fetch Repository
         uses: actions/checkout@v4
@@ -29,9 +38,34 @@ jobs:
         with:
           go-version: ${{ matrix.go-version }}
 
+      # Whole-suite retry, without -race and without -shuffle in the blocking
+      # job: the suite has pre-existing data races (whose -race abort fails the
+      # run) and order-dependent tests (which -shuffle exposes). gotestsum's
+      # per-test rerun cannot absorb these because the recover test prints a
+      # literal "panic:" (via debug.Stack), which makes gotestsum treat the
+      # whole run as a suspected panic and refuse to rerun. Race detection runs
+      # in the separate non-blocking Race job below.
       - name: Run Test
         uses: nick-fields/retry@v2
         with:
           max_attempts: 3
           timeout_minutes: 15
-          command: go test ./... -v -race -count=1
+          command: go test ./... -count=1
+
+  # Non-blocking: surfaces the suite's pre-existing data races without gating
+  # merges on them. Fix the races over time, then fold -race into Build.
+  Race:
+    runs-on: ubuntu-latest
+    continue-on-error: true
+    timeout-minutes: 20
+    steps:
+      - name: Fetch Repository
+        uses: actions/checkout@v4
+
+      - name: Install Go
+        uses: actions/setup-go@v6
+        with:
+          go-version: 1.23.x
+
+      - name: Test (race)
+        run: go test ./... -race -count=1 -shuffle=on