diff --git a/.github/workflows/benchmark.yaml b/.github/workflows/benchmark.yaml
index e18b7fa9aa..7d48399daa 100644
--- a/.github/workflows/benchmark.yaml
+++ b/.github/workflows/benchmark.yaml
@@ -15,7 +15,7 @@ jobs:
   benchmarks:
     name: "Run Full Go benchmarks on Main"
     runs-on: "depot-ubuntu-24.04-arm-8" # same as build-test.yaml
-    timeout-minutes: 15
+    timeout-minutes: 20
     steps:
       - uses: "actions/checkout@v7"
         with:
diff --git a/internal/datastore/benchmark/driver_bench_test.go b/internal/datastore/benchmark/driver_bench_test.go
index 0255c68fe4..8bead1aaa1 100644
--- a/internal/datastore/benchmark/driver_bench_test.go
+++ b/internal/datastore/benchmark/driver_bench_test.go
@@ -57,6 +57,10 @@ var sortOrders = map[string]options.SortOrder{
 }
 
 func BenchmarkDatastoreDriver(b *testing.B) {
+	if testing.Short() {
+		b.Skip("skipping datastore driver benchmarks in -short mode")
+	}
+
 	for _, driver := range drivers {
 		b.Run(driver.name+driver.suffix, func(b *testing.B) {
 			b.StopTimer()
diff --git a/internal/fdw/README.md b/internal/fdw/README.md
index af2465bff7..28e07aa326 100644
--- a/internal/fdw/README.md
+++ b/internal/fdw/README.md
@@ -226,8 +226,8 @@ SPICEDB_SHUTDOWN_GRACE_PERIOD
 | `subject_type` | text | Subject type |
 | `subject_id` | text | Subject ID |
 | `optional_subject_relation` | text | Optional subject relation |
-| `optional_caveat_name` | text | Optional caveat name |
-| `optional_caveat_context` | jsonb | Optional caveat context |
+| `caveat_name` | text | Optional caveat name |
+| `caveat_context` | text | Optional caveat context |
 | `consistency` | text | Consistency token (ZedToken) |
 
 **Supported Operations:**
diff --git a/internal/services/integrationtesting/benchmark_executors_test.go b/internal/services/integrationtesting/benchmark_executors_test.go
index 1a6738c20d..9c9aa2cbf8 100644
--- a/internal/services/integrationtesting/benchmark_executors_test.go
+++ b/internal/services/integrationtesting/benchmark_executors_test.go
@@ -55,7 +55,7 @@ const classicDispatchDepthRemaining = 50
 func BenchmarkExecutors(b *testing.B) {
 	for _, engineID := range enginesToBenchmark {
 		if testing.Short() && engineID != "memory" {
-			continue
+			b.Skip("skipping non-memory driver benchmarks in -short mode")
 		}
 		b.Run(engineID, func(b *testing.B) {
 			for _, scenarioName := range executorScenarios {
diff --git a/internal/services/integrationtesting/benchmark_test.go b/internal/services/integrationtesting/benchmark_test.go
index 2ff1277619..7bc480a6a0 100644
--- a/internal/services/integrationtesting/benchmark_test.go
+++ b/internal/services/integrationtesting/benchmark_test.go
@@ -299,10 +299,8 @@ var allScenarios = []benchmarkScenario{
 // Hierarchy: engine → scenario (test data) → operation → {dispatch, queryplan}
 func BenchmarkServices(b *testing.B) {
 	for _, engineID := range enginesToBenchmark {
-		if testing.Short() {
-			if engineID != "memory" {
-				continue
-			}
+		if testing.Short() && engineID != "memory" {
+			b.Skip("skipping non-memory driver benchmarks in -short mode")
 		}
 		b.Run(engineID, func(b *testing.B) {
 			for _, scenario := range allScenarios {
diff --git a/magefiles/benchmark.go b/magefiles/benchmark.go
index 4bb0c6664e..e224cd437f 100644
--- a/magefiles/benchmark.go
+++ b/magefiles/benchmark.go
@@ -3,6 +3,8 @@
 package main
 
 import (
+	"fmt"
+
 	"github.com/magefile/mage/mg"
 	"github.com/magefile/mage/sh"
 )
@@ -11,10 +13,29 @@ type Benchmark mg.Namespace
 
 // All Runs all benchmarks
 func (b Benchmark) All() error {
-	return sh.RunV("go", "test", "./...", "-ldflags=-checklinkname=0", "-bench", ".", "-benchtime", "5s", "-tags", "memoryprotection", "-timeout", "0", "-run=XXX", "-cpu", "1", "-benchmem")
+	if err := prebuildTestBinaries(); err != nil {
+		return err
+	}
+	fmt.Println("running all benchmarks")
+	// -p 1 runs one package's test binary at a time so no other package runs during a measurement window.
+	// -cpu 1 sets GOMAXPROCS=1 inside each benchmark, measuring single-threaded per-op latency
+	return sh.RunV("go", "test", "-p", "1", "./...", "-ldflags=-checklinkname=0", "-bench", ".", "-benchtime", "5s", "-tags", "memoryprotection", "-timeout", "0", "-run=XXX", "-cpu", "1", "-benchmem")
 }
 
 // Short runs the short benchmarks for individual PRs on CI -- less time, only the short ones
 func (b Benchmark) Short() error {
-	return sh.RunV("go", "test", "./...", "-ldflags=-checklinkname=0", "-bench", ".", "-benchtime", "2s", "-tags", "memoryprotection", "-timeout", "0", "-run=XXX", "-cpu", "1", "-benchmem", "-short")
+	if err := prebuildTestBinaries(); err != nil {
+		return err
+	}
+	fmt.Println("running short benchmarks")
+	// -p 1 runs one package's test binary at a time so no other package runs during a measurement window.
+	// -cpu 1 sets GOMAXPROCS=1 inside each benchmark, measuring single-threaded per-op latency
+	return sh.RunV("go", "test", "-p", "1", "./...", "-ldflags=-checklinkname=0", "-bench", ".", "-benchtime", "2s", "-tags", "memoryprotection", "-timeout", "0", "-run=XXX", "-cpu", "1", "-benchmem", "-short")
+}
+
+// prebuildTestBinaries compiles and links every test binary in parallel (-run=XXX runs nothing),
+// with the same tags/ldflags as the benchmark runs so they hit the build cache.
+func prebuildTestBinaries() error {
+	fmt.Println("prebuilding all test binaries in parallel")
+	return sh.RunV("go", "test", "./...", "-ldflags=-checklinkname=0", "-tags", "memoryprotection", "-timeout", "0", "-run=XXX")
 }