diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index ed4045e..fdb5b12 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -23,7 +23,7 @@ jobs:
       - name: Setup Go
         uses: actions/setup-go@v5
         with:
-          go-version: '1.25.x'
+          go-version: '1.26.x'
           cache: true
 
       - name: Verify formatting
@@ -51,7 +51,7 @@ jobs:
       - name: Setup Go
         uses: actions/setup-go@v5
         with:
-          go-version: '1.25.x'
+          go-version: '1.26.x'
           cache: true
 
       - name: Test with coverage (exclude internal/*)
@@ -118,7 +118,7 @@ jobs:
       - name: Setup Go
         uses: actions/setup-go@v5
         with:
-          go-version: '1.25.x'
+          go-version: '1.26.x'
           cache: true
 
       - name: Cross compile (packages + tests, no execution)
diff --git a/go.mod b/go.mod
index 5279650..d2380e3 100644
--- a/go.mod
+++ b/go.mod
@@ -1,3 +1,3 @@
 module code.hybscloud.com/spin
 
-go 1.25
+go 1.26
diff --git a/lock.go b/lock.go
index da59319..cc33fd0 100644
--- a/lock.go
+++ b/lock.go
@@ -12,22 +12,44 @@ import (
 // Lock is a minimal, non-fair spin lock intended for very short
 // critical sections on hot paths. It avoids allocations and OS mutex
 // overhead but should not be used as a general-purpose lock.
+//
+// Acquisition uses FAA (Fetch-And-Add) with a TTAS (test-and-test-and-set)
+// slow path. FAA completes in a single atomic instruction (LOCK XADD on
+// x86, LDADDAL on arm64) — under contention, CAS produces O(n²) cache-line
+// invalidations while FAA keeps traffic at O(n). The TTAS slow path spins
+// on read-only Load (MESI Shared) to avoid bouncing the cache line, and
+// only attempts FAA when the lock appears free.
 type Lock struct {
 	_ noCopy
 	n atomic.Uintptr
 }
 
-// Lock acquires the lock, spinning with an adaptive backoff.
+// Lock acquires the lock.
+// Fast path: single FAA. Slow path: TTAS (test-and-test-and-set) —
+// spin on read-only Load to keep the cache line in MESI Shared state,
+// then attempt FAA only when the lock appears free.
 func (sl *Lock) Lock() {
-	for {
-		n := sl.n.Add(1)
-		if n == 1 {
-			return
-		} else if n < 4 {
+	if sl.n.Add(1) == 1 {
+		return
+	}
+	sl.lockSlow()
+}
+
+func (sl *Lock) lockSlow() {
+	for i := 0; ; i++ {
+		// Spin on Load (read-only): multiple cores hold the line in
+		// Shared state with zero cross-core invalidations. Only
+		// transition to Add (write) when the lock appears released.
+		if sl.n.Load() == 0 {
+			if sl.n.Add(1) == 1 {
+				return
+			}
+		}
+		if i >= 4 {
+			runtime.Gosched()
+		} else {
 			Pause(defaultPauseCycles)
-			continue
 		}
-		runtime.Gosched()
 	}
 }
 
@@ -38,6 +60,16 @@ func (sl *Lock) Unlock() {
 
 // Try attempts to acquire the lock without blocking.
 // It returns true if the lock was acquired.
+//
+// Uses FAA rather than CAS: on x86 LOCK CMPXCHG takes exclusive
+// cache-line ownership regardless of comparison outcome, so a failed
+// CAS costs the same coherence traffic as FAA. FAA keeps the atomic
+// pattern consistent across the Lock API.
+//
+// A failed Try increments n without a corresponding decrement.
+// This is intentional: Unlock resets n to zero via Store(0),
+// clearing all accumulated increments. The uintptr counter space
+// is large enough that overflow is not a practical concern.
 func (sl *Lock) Try() bool {
 	return sl.n.Add(1) == 1
 }
diff --git a/pause_test.go b/pause_test.go
index 818d359..815a087 100644
--- a/pause_test.go
+++ b/pause_test.go
@@ -15,7 +15,7 @@ import (
 )
 
 func TestPause(t *testing.T) {
-	// Default (20 cycles)
+	// Default (30 cycles)
 	spin.Pause()
 
 	// Single cycle
diff --git a/yield.go b/yield.go
index f8ef4d5..0017e52 100644
--- a/yield.go
+++ b/yield.go
@@ -9,7 +9,10 @@ import (
 	"time"
 )
 
-var yieldDuration = 250 * time.Microsecond
+var (
+	yieldDuration = 250 * time.Microsecond
+	yieldMu       Lock // protects yieldDuration
+)
 
 // Yield cooperatively yields execution to reduce CPU burn in tight loops.
 //
@@ -22,7 +25,9 @@ var yieldDuration = 250 * time.Microsecond
 //
 // For automatic adaptive backoff in tight loops, use Wait instead.
 func Yield(duration ...time.Duration) {
+	yieldMu.Lock()
 	d := yieldDuration
+	yieldMu.Unlock()
 	if len(duration) > 0 {
 		d = max(0, duration[0])
 	}
@@ -34,10 +39,13 @@ func Yield(duration ...time.Duration) {
 }
 
 // SetYieldDuration sets the base duration unit for Yield().
+// Safe for concurrent use.
 // Recommended: 50-250 microseconds for real-time systems, 1-4 ms for general workloads.
 func SetYieldDuration(d time.Duration) {
 	if d < 0 {
 		d = 0
 	}
+	yieldMu.Lock()
 	yieldDuration = d
+	yieldMu.Unlock()
 }