fix(qwen36): keep CUDA tier heat adaptive

c/qwen36_tier.c

@@ -6,6 +6,7 @@
 #include <pthread.h>
 #include "qwen36_tier.h"
 #include "backend_cuda.h"
+#include "tier.h"
 
 #define QT_MAX_DEV 8
 #define QT_QCAP 48            /* upload queue depth (staging ~1.6 MB/entry) */
@@ -330,11 +331,15 @@ void qt_fill_wait(void){
     pthread_mutex_unlock(&G.mx);
 }
 
-/* LFRU swap check (every 16 ticks = tokens): per device, coldest resident vs
- * hottest non-resident, with the tier.h hysteresis. */
+/* Adaptive swap check (every 16 ticks = tokens): per device, coldest resident
+ * vs hottest non-resident. Decay every 1024 ticks so an old workload cannot
+ * permanently own the tier; admission uses the shared tier.h contract. */
 static void qt_lfru_tick_locked(void){
-    if(++G.tick % 16) return;
     size_t n=(size_t)G.nl*G.ne;
+    G.tick++;
+    if(!(G.tick%1024))
+        for(size_t i=0;i<n;i++) G.slot[i].heat=tier_decay_value(G.slot[i].heat);
+    if(G.tick%16) return;
     for(int di=0;di<G.ndev;di++){
         int cold=-1, hot=-1; uint32_t ch=0, hh=0;
         for(size_t i=0;i<n;i++){
@@ -345,7 +350,7 @@ static void qt_lfru_tick_locked(void){
             else if(!s->resident && !s->queued && s->g4){ if(hot<0||s->heat>hh){ hot=(int)i; hh=s->heat; } }
         }
         if(cold<0||hot<0) continue;
-        if(hh<=ch+(ch>>2)+4) continue;                    /* hysteresis as in tier.h */
+        if(!tier_should_promote(hh,ch)) continue;
         QSlot *v=&G.slot[cold];
         v->resident=0;                                    /* CPU fallback from now on */
         if(enqueue_locked(hot/G.ne,hot%G.ne,cold/G.ne,cold%G.ne,0)) G.swaps++;

c/tests/test_tier.c

@@ -7,6 +7,11 @@ static int fail(const char *message){
 }
 
 int main(void){
+    if(!tier_should_promote(130,100)) return fail("shared hysteresis admits hot expert");
+    if(tier_should_promote(129,100)) return fail("shared hysteresis blocks marginal expert");
+    if(tier_should_promote(UINT32_MAX-1,UINT32_MAX))
+        return fail("saturated hysteresis threshold wrapped");
+    if(tier_decay_value(9)!=4) return fail("shared heat decay");
     uint32_t heat[6]={20,2,8,3,30,1};
     int pinned[2]={0,1}, slot=-1, eid=-1; long gain=0;
     if(!tier_pick_swap(heat,6,pinned,2,&slot,&eid,&gain)) return fail("hot expert not promoted");

c/tier.h

@@ -3,6 +3,16 @@
 
 #include <stdint.h>
 
+/* Shared admission contract for every adaptive resident tier.  Widen before
+ * adding the margin: a saturated uint32 heat counter must become sticky, not
+ * wrap the threshold and admit a colder expert. */
+static int tier_should_promote(uint32_t hot, uint32_t cold){
+    uint64_t threshold=(uint64_t)cold+((uint64_t)cold>>2)+4u;
+    return (uint64_t)hot>threshold;
+}
+
+static uint32_t tier_decay_value(uint32_t heat){ return heat>>1; }
+
 /* Pick one RAM/VRAM hot-store slot to replace from recent routing heat.
  * The fixed margin handles tiny samples; the 25% margin prevents ping-pong. */
 static int tier_pick_swap(const uint32_t *heat, int nexpert,
@@ -19,7 +29,7 @@ static int tier_pick_swap(const uint32_t *heat, int nexpert,
     }
     if(hot<0) return 0;
     uint32_t fc=heat[pinned[cold]];
-    if(fh<=fc+(fc>>2)+4) return 0;
+    if(!tier_should_promote(fh,fc)) return 0;
     *slot=cold; *eid=hot; *gain=(long)fh-(long)fc;
     return 1;
 }
@@ -54,7 +64,7 @@ static int tier_pick_lfru(const uint32_t *heat, const uint32_t *last, uint32_t c
 }
 
 static void tier_decay(uint32_t *heat, int nexpert){
-    for(int e=0;e<nexpert;e++) heat[e]>>=1;
+    for(int e=0;e<nexpert;e++) heat[e]=tier_decay_value(heat[e]);
 }
 
 #endif

docs/qwen36-cuda-tier.md

@@ -10,7 +10,9 @@ more GPUs and computed there through the existing shared CUDA backend
 
 - **Home device:** expert `eid` lives on GPU `eid % n_gpus`; no duplicates.
 - **Placement:** routing heat decides who earns VRAM (LFRU semantics from
-  `tier.h`, 25%+4 hysteresis). A parallel **warmstart** fills the per-device
+  `tier.h`, 25%+4 hysteresis). Runtime heat is halved every 1024 decode ticks,
+  so a long-lived process can replace experts from an old workload instead of
+  permanently freezing its initial hot set. A parallel **warmstart** fills the per-device
   budget before the first token — ordered by a persisted heat table
   (`HEAT_FILE`) when present, so a second run starts fully placed.
 - **Decode:** per (token, layer) the resident experts are issued as async