diff --git a/apps/desktop/electron/main.ts b/apps/desktop/electron/main.ts
index 0e5ec73363f32..f124456f56696 100644
--- a/apps/desktop/electron/main.ts
+++ b/apps/desktop/electron/main.ts
@@ -280,7 +280,7 @@ import {
   localRouteFallbackProfiles,
   undialedSshRouteSeeds
 } from './plugin-profile-routes'
-import { selectPoolEvictions } from './pool-eviction'
+import { selectPoolEvictions, selectSlotDisplacementVictim } from './pool-eviction'
 import { clampPoolLimits, parsePoolLimits, POOL_LIMITS_DEFAULTS } from './pool-limits'
 import {
   LocalBackendSpawnCoordinator,
@@ -12249,6 +12249,31 @@ function evictLruPoolBackends(keep) {
   }
 }
 
+// Demand displacement for the hard spawn cap (#102163). evictLruPoolBackends()
+// only evicts backends idle past the keepalive window, so a fast A→B→C→D
+// switch (every holder still fresh) converges nothing and the 4th spawn queued
+// behind full slots until its 30s ticket expired. A real profile switch
+// outranks background pressure: stop the stalest RUNNING backend and wait for
+// its slot, so the coordinator hands it to the waiter below instead of timing
+// out. Starting/queued entries are never victims (preempting a mid-boot spawn
+// just moves the error to an earlier click); with no running victim this is a
+// no-op and genuine capacity pressure still surfaces the queue timeout.
+// ponytail: single displacement per spawn; a many-way concurrent boot storm
+// with zero running backends still funnels through the queue timeout — widen
+// to a drain loop only if that path is ever observed in the wild.
+async function displaceStalestPoolBackendForSlot(excludeKey) {
+  const victim = selectSlotDisplacementVictim(backendPool.entries(), excludeKey)
+
+  if (victim === null) {
+    return false
+  }
+
+  rememberLog(`Profile backend "${excludeKey}" displacing stalest pool backend "${victim}" for a free local slot`)
+  await stopPoolBackend(victim)
+
+  return true
+}
+
 function startPoolIdleReaper() {
   if (poolIdleReaper) {
     return
@@ -12386,6 +12411,9 @@ async function spawnPoolBackend(profile, entry, opts: { forceLocal?: boolean; po
     rememberLog(
       `Profile backend "${profile}" waiting for a free local slot (${localBackendSpawnCoordinator.activeCount}/${poolMaxBackends()} busy, ${localBackendSpawnCoordinator.queuedCount} queued)`
     )
+    // Demand displacement (#102163): converge the pool now instead of riding
+    // the queue to its timeout — see displaceStalestPoolBackendForSlot.
+    await displaceStalestPoolBackendForSlot(poolKey)
   }
 
   entry.releaseLocalBackendSlot = await spawnRequest.acquired
diff --git a/apps/desktop/electron/pool-eviction.test.ts b/apps/desktop/electron/pool-eviction.test.ts
index 4ad3c3fc0def3..073de6b93cbeb 100644
--- a/apps/desktop/electron/pool-eviction.test.ts
+++ b/apps/desktop/electron/pool-eviction.test.ts
@@ -151,3 +151,14 @@ test('#95189: a backend genuinely idle for minutes IS evicted (#95189 long-windo
   // keep=1, idle is over the cap AND past the fresh window → evicted.
   assert.deepEqual(selectPoolEvictions(entries, 1, NOW, FRESH_MS), ['idle'])
 })
+
+test('slot displacement picks stalest running backend, never self or starters', async () => {
+  const { selectSlotDisplacementVictim } = await import('./pool-eviction')
+  const entries: [string, ReturnType<typeof spawned>][] = [
+    ['a', spawned(9_000)],
+    ['b', spawned(3_000)],
+    ['d', spawned(0)]
+  ]
+  assert.equal(selectSlotDisplacementVictim(entries, 'd'), 'a')
+  assert.equal(selectSlotDisplacementVictim([['d', spawned(0)]], 'd'), null)
+})
diff --git a/apps/desktop/electron/pool-eviction.ts b/apps/desktop/electron/pool-eviction.ts
index 628a3311993db..cc7a48b1c8700 100644
--- a/apps/desktop/electron/pool-eviction.ts
+++ b/apps/desktop/electron/pool-eviction.ts
@@ -56,3 +56,38 @@ export function selectPoolEvictions<K>(
 
   return evictions
 }
+
+/**
+ * Demand displacement victim for the hard local-backend spawn cap (#102163).
+ *
+ * selectPoolEvictions() spares keepalive-fresh backends, so a fast A→B→C→D
+ * profile switch converges nothing and the 4th spawn queued behind full
+ * slots until its ticket expired ("timed out while waiting for a free
+ * slot"). A real profile switch outranks background pressure: displace the
+ * stalest RUNNING backend (its session persists; reopening respawns) so the
+ * pool converges instead of erroring. Only entries holding a child process
+ * free a coordinator slot — starting/queued/descriptor entries are never
+ * victims, nor is the incoming spawn itself (`exclude`).
+ */
+export function selectSlotDisplacementVictim<K>(
+  entries: Iterable<[K, PoolEvictionEntry]>,
+  exclude: K
+): K | null {
+  let victim: K | null = null
+  let victimAt = Number.POSITIVE_INFINITY
+
+  for (const [key, entry] of entries) {
+    if (key === exclude || !entry.process) {
+      continue
+    }
+
+    const at = typeof entry.lastActiveAt === 'number' ? entry.lastActiveAt : 0
+
+    if (at < victimAt) {
+      victim = key
+      victimAt = at
+    }
+  }
+
+  return victim
+}