P7: delete the dead placement machinery

skainet-backends/skainet-backend-cpu/api/jvm/skainet-backend-cpu.api

@@ -17,7 +17,6 @@ public final class sk/ainet/context/DirectCpuExecutionContext : sk/ainet/context
 	public fun getHooks ()Lsk/ainet/lang/nn/hooks/ForwardHooks;
 	public fun getInTraining ()Z
 	public fun getMemoryInfo ()Lsk/ainet/context/MemoryInfo;
-	public fun getMemoryPlanner ()Lsk/ainet/lang/tensor/storage/MemoryPlanner;
 	public fun getMemoryScope ()Lsk/ainet/lang/memory/Scope;
 	public fun getMemoryTracker ()Lsk/ainet/lang/tensor/storage/MemoryTracker;
 	public fun getObservers ()Lsk/ainet/context/ExecutionObserverRegistry;

skainet-compile/skainet-compile-dag/api/jvm/skainet-compile-dag.api

@@ -172,7 +172,6 @@ public final class sk/ainet/lang/graph/DefaultGraphExecutionContext : sk/ainet/l
 	public fun getHooks ()Lsk/ainet/lang/nn/hooks/ForwardHooks;
 	public fun getInTraining ()Z
 	public fun getMemoryInfo ()Lsk/ainet/context/MemoryInfo;
-	public fun getMemoryPlanner ()Lsk/ainet/lang/tensor/storage/MemoryPlanner;
 	public fun getMemoryScope ()Lsk/ainet/lang/memory/Scope;
 	public fun getMemoryTracker ()Lsk/ainet/lang/tensor/storage/MemoryTracker;
 	public fun getObservers ()Lsk/ainet/context/ExecutionObserverRegistry;
@@ -449,7 +448,6 @@ public final class sk/ainet/lang/graph/exec/GraphExecutionContext$DefaultImpls {
 	public static fun full (Lsk/ainet/lang/graph/exec/GraphExecutionContext;Lsk/ainet/lang/tensor/Shape;Lkotlin/reflect/KClass;Ljava/lang/Number;)Lsk/ainet/lang/tensor/Tensor;
 	public static fun getHooks (Lsk/ainet/lang/graph/exec/GraphExecutionContext;)Lsk/ainet/lang/nn/hooks/ForwardHooks;
 	public static fun getInTraining (Lsk/ainet/lang/graph/exec/GraphExecutionContext;)Z
-	public static fun getMemoryPlanner (Lsk/ainet/lang/graph/exec/GraphExecutionContext;)Lsk/ainet/lang/tensor/storage/MemoryPlanner;
 	public static fun getMemoryScope (Lsk/ainet/lang/graph/exec/GraphExecutionContext;)Lsk/ainet/lang/memory/Scope;
 	public static fun getMemoryTracker (Lsk/ainet/lang/graph/exec/GraphExecutionContext;)Lsk/ainet/lang/tensor/storage/MemoryTracker;
 	public static fun getScratch (Lsk/ainet/lang/graph/exec/GraphExecutionContext;)Lsk/ainet/lang/tensor/scratch/ScratchPool;

skainet-io/skainet-io-safetensors/src/commonTest/kotlin/sk/ainet/io/safetensors/StorageAwareSafeTensorsLoaderTest.kt

@@ -6,7 +6,6 @@ import sk.ainet.io.RandomAccessSource
 import sk.ainet.lang.tensor.storage.LogicalDType
 import sk.ainet.lang.tensor.storage.MemoryDomain
 import sk.ainet.lang.tensor.storage.Ownership
-import sk.ainet.lang.tensor.storage.Residency
 import kotlin.test.Test
 import kotlin.test.assertEquals
 import kotlin.test.assertFalse
@@ -127,7 +126,6 @@ class StorageAwareSafeTensorsLoaderTest {
         assertTrue(storage.isFileBacked)
         assertEquals(Ownership.FILE_BACKED, storage.ownership)
         assertEquals(MemoryDomain.MMAP_FILE, storage.placement.domain)
-        assertEquals(Residency.PERSISTENT, storage.placement.residency)
         assertFalse(storage.isMutable)
     }
 

skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/context/ExecutionContext.kt

@@ -8,7 +8,6 @@ import sk.ainet.lang.tensor.operators.OpsBoundTensor
 import sk.ainet.lang.tensor.ops.TensorOps
 import sk.ainet.lang.tensor.scratch.NoopScratchPool
 import sk.ainet.lang.tensor.scratch.ScratchPool
-import sk.ainet.lang.tensor.storage.MemoryPlanner
 import sk.ainet.lang.tensor.storage.MemoryTracker
 import sk.ainet.lang.types.DType
 import kotlin.reflect.KClass
@@ -187,9 +186,6 @@ public interface ExecutionContext {
     public val memoryInfo: MemoryInfo
     public val executionStats: ExecutionStats
 
-    /** Memory planner for resolving placement intents. Default: CPU-only. */
-    public val memoryPlanner: MemoryPlanner get() = MemoryPlanner()
-
     /** Memory tracker for observability and copy tracing. Default: no-op (not tracking). */
     public val memoryTracker: MemoryTracker? get() = null
 }

skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/tensor/ops/turboquant/TurboQuantPresets.kt

@@ -2,7 +2,6 @@ package sk.ainet.lang.tensor.ops.turboquant
 
 import sk.ainet.lang.tensor.storage.KvCacheConfig
 import sk.ainet.lang.tensor.storage.Placement
-import sk.ainet.lang.tensor.storage.Residency
 import sk.ainet.lang.tensor.storage.TensorEncoding
 
 /**
@@ -39,7 +38,7 @@ public object TurboQuantPresets {
             maxSeqLen = maxSeqLen,
             keyEncoding = TensorEncoding.Q8_0,
             valueEncoding = TensorEncoding.TurboQuantPolar(bitsPerElement = 4),
-            placement = Placement.CPU_HEAP.copy(residency = Residency.PERSISTENT)
+            placement = Placement.CPU_HEAP
         ),
         keyQuantConfig = null, // Q8_0 uses standard quantization, not TurboQuant
         valueQuantConfig = TurboQuantConfig.polarOnly(bits = 4)
@@ -65,7 +64,7 @@ public object TurboQuantPresets {
             maxSeqLen = maxSeqLen,
             keyEncoding = TensorEncoding.TurboQuantPolar(bitsPerElement = 4),
             valueEncoding = TensorEncoding.TurboQuantPolar(bitsPerElement = 4),
-            placement = Placement.CPU_HEAP.copy(residency = Residency.PERSISTENT)
+            placement = Placement.CPU_HEAP
         ),
         keyQuantConfig = TurboQuantConfig.polarOnly(bits = 4),
         valueQuantConfig = TurboQuantConfig.polarOnly(bits = 4)
@@ -92,7 +91,7 @@ public object TurboQuantPresets {
             maxSeqLen = maxSeqLen,
             keyEncoding = TensorEncoding.TurboQuantPolar(bitsPerElement = 3),
             valueEncoding = TensorEncoding.TurboQuantPolar(bitsPerElement = 3),
-            placement = Placement.CPU_HEAP.copy(residency = Residency.PERSISTENT)
+            placement = Placement.CPU_HEAP
         ),
         keyQuantConfig = TurboQuantConfig.polarOnly(bits = 3),
         valueQuantConfig = TurboQuantConfig.polarOnly(bits = 3)

skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/tensor/storage/PlacementAnnotations.kt → skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/tensor/storage/KvCacheAnnotations.kt

@@ -1,51 +1,5 @@
 package sk.ainet.lang.tensor.storage
 
-/**
- * Declares placement intent for a tensor parameter or property.
- *
- * The [MemoryPlanner] reads these annotations (via reflection or codegen)
- * to decide where tensors should be allocated. This expresses *intent*,
- * not a hard guarantee — the planner may fall back if the target is
- * unavailable and [requirement] is [Requirement.PREFERRED].
- *
- * Example:
- * ```kotlin
- * @Place(device = DeviceKind.GPU, memory = MemoryDomain.DEVICE_LOCAL)
- * val projectionWeight: Tensor<FP32, Float>
- * ```
- */
-@Target(AnnotationTarget.PROPERTY, AnnotationTarget.VALUE_PARAMETER, AnnotationTarget.FIELD)
-@Retention(AnnotationRetention.RUNTIME)
-public annotation class Place(
-    val device: DeviceKind = DeviceKind.AUTO,
-    val memory: MemoryDomain = MemoryDomain.HOST_HEAP,
-    val requirement: Requirement = Requirement.PREFERRED
-)
-
-/**
- * Marks a tensor as an immutable weight that should be file-backed
- * (memory-mapped) when possible.
- *
- * Equivalent to `@Place(device = CPU, memory = MMAP_FILE)` with
- * [Residency.PERSISTENT]. The planner treats these tensors as
- * read-only and long-lived, preferring OS-paged file access over
- * heap allocation.
- *
- * Example:
- * ```kotlin
- * @Weights
- * val embeddings: Tensor<FP32, Float>
- *
- * @Weights(memory = MemoryDomain.HOST_HEAP)  // force heap for small weights
- * val biasVector: Tensor<FP32, Float>
- * ```
- */
-@Target(AnnotationTarget.PROPERTY, AnnotationTarget.VALUE_PARAMETER, AnnotationTarget.FIELD)
-@Retention(AnnotationRetention.RUNTIME)
-public annotation class Weights(
-    val memory: MemoryDomain = MemoryDomain.MMAP_FILE
-)
-
 /**
  * Configures TurboQuant KV-cache compression for an attention layer.
  *

skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/tensor/storage/KvCacheStore.kt

@@ -300,7 +300,7 @@ public data class KvCacheConfig(
     val maxSeqLen: Int,
     val keyEncoding: TensorEncoding = TensorEncoding.Dense(4),
     val valueEncoding: TensorEncoding = TensorEncoding.Dense(4),
-    val placement: Placement = Placement.CPU_HEAP.copy(residency = Residency.PERSISTENT),
+    val placement: Placement = Placement.CPU_HEAP,
     /**
      * The dtype the key ring stores; with [keyEncoding] it forms the store's `keyFormat` (#1077).
      * `FP32` is what the dense store has always held; `BF16`/`FP16` halve the ring.

skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/tensor/storage/Placement.kt

@@ -4,15 +4,16 @@ package sk.ainet.lang.tensor.storage
  * High-level placement descriptor: where a tensor lives and how the runtime
  * should manage it.
  *
- * Placement is *intent* — it tells the planner what to aim for but does not
- * encode backend scratch-memory details. The planner resolves placement to
- * a concrete [BufferHandle] and falls back if the preferred target is
- * unavailable.
+ * Placement is *intent* — it tells the runtime what to aim for but does not
+ * encode backend scratch-memory details.
+ *
+ * Lifetime is deliberately not part of placement: how long bytes live is a
+ * scope decision (`sk.ainet.lang.memory.ScopeKind`), and how a weight is
+ * staged at load is a resolver decision (`WeightForm.WeightResidency`).
  */
 public data class Placement(
     val device: DeviceKind = DeviceKind.CPU,
     val domain: MemoryDomain = MemoryDomain.HOST_HEAP,
-    val residency: Residency = Residency.PERSISTENT,
     val requirement: Requirement = Requirement.PREFERRED,
     val fallback: DeviceKind = DeviceKind.CPU
 ) {
@@ -21,23 +22,20 @@ public data class Placement(
         public val CPU_HEAP: Placement = Placement(
             device = DeviceKind.CPU,
             domain = MemoryDomain.HOST_HEAP,
-            residency = Residency.TRANSIENT,
             requirement = Requirement.PREFERRED
         )
 
         /** File-backed placement for immutable model weights. */
         public val MMAP_WEIGHTS: Placement = Placement(
             device = DeviceKind.CPU,
             domain = MemoryDomain.MMAP_FILE,
-            residency = Residency.PERSISTENT,
             requirement = Requirement.PREFERRED
         )
 
         /** GPU-preferred placement with CPU fallback. */
         public val GPU_PREFERRED: Placement = Placement(
             device = DeviceKind.GPU,
             domain = MemoryDomain.DEVICE_LOCAL,
-            residency = Residency.PERSISTENT,
             requirement = Requirement.PREFERRED,
             fallback = DeviceKind.CPU
         )
@@ -66,13 +64,6 @@ public enum class MemoryDomain {
     DEVICE_LOCAL
 }
 
-public enum class Residency {
-    /** Short-lived: activations, temporaries, intermediate results. */
-    TRANSIENT,
-    /** Long-lived: model weights, embeddings, caches. */
-    PERSISTENT
-}
-
 public enum class Requirement {
     /** Best-effort: fall back to [Placement.fallback] if unavailable. */
     PREFERRED,