test: stabilize platform-sensitive CI checks (#1303)

The Linux race suite inherits Go’s ten-minute default package timeout
even though the ordinary suite allows twenty minutes. On the failed main
run, `internal/sync` reached that boundary without reporting a data
race, turning an instrumentation-slow package into a false CI failure.

Windows hit a similar timing assumption: vector serving stayed open
correctly while the API build was blocked, but the test allowed only two
seconds for index finalization after release. This keeps both lifecycle
assertions while giving finalization a bounded allowance appropriate for
loaded Windows runners.

This is intentionally a reliability fix, not a suite-sharding change.
`internal/sync` and `internal/artifact` are already near co-critical
paths; the larger follow-up opportunity is to run the separate non-race
DuckDB command in parallel with the race suite.

Author: mariusvniekerk

.github/workflows/ci.yml

@@ -176,7 +176,7 @@ jobs:
         if: runner.os == 'Linux'
         run: |
           go test -tags "fts5" ./internal/duckdb -count=1
-          go test -tags "fts5" -race $(go list ./... | grep -v '/internal/duckdb$') -count=1
+          go test -tags "fts5" -race $(go list ./... | grep -v '/internal/duckdb$') -count=1 -timeout=20m
         env:
           CGO_ENABLED: "1"
 

cmd/agentsview/embed_scheduler_test.go

@@ -1522,7 +1522,7 @@ func TestVectorServingCloseWaitsForAPIStartedRecallBuild(t *testing.T) {
 	select {
 	case closeErr := <-closeDone:
 		require.NoError(t, closeErr)
-	case <-time.After(2 * time.Second):
+	case <-time.After(10 * time.Second):
 		require.Fail(t, "vector serving did not close after the API build completed")
 	}
 }