Add explicit CPU-state AdamW with scoped CUDA resume evidence

docs/CPU_STATE_OPTIMIZER.md

@@ -0,0 +1,76 @@
+# Explicit CPU optimizer state
+
+`cpu_state_streaming_v1` is an experimental, explicitly selected CUDA AdamW
+backend. It requires PyTorch `2.10.0+cu128` and trainable CUDA FP32 parameters,
+including FP32 parameters used under BF16 autocast. Default and fused optimizer
+selection and their recorded configuration hashes are preserved.
+
+Select it in the training configuration with an explicit persistent state budget:
+
+```json
+{
+  "training": {
+    "optimizer_backend": "cpu_state_streaming_v1",
+    "optimizer_state_max_bytes": 4000000000
+  }
+}
+```
+
+The budget bounds two CPU moment tensors and one FP32 CPU scalar step per
+trainable parameter. It does not bound total host RAM, checkpoint deserialization,
+model parameters or gradients. The backend rejects an insufficient budget before
+construction. Each update stages one parameter's moments on CUDA and calls the
+pinned stock single-tensor AdamW functional. Global gradient clipping remains in
+the trainer. There is no automatic fallback or allocator-cap increase.
+
+Algorithm version and ordered trainable parameter names, shapes and dtypes enter
+the opt-in configuration identity. Resume across default, fused and CPU-state
+backends is refused. State loading validates the complete payload before changing
+the optimizer, keeps moments on CPU, and rejects malformed or aliased state.
+Live parameter replacement or reordering is also refused. The trainer releases
+temporary checkpoint/model objects after restoring optimizer, scheduler, sampler
+and RNG state; the CLI independently releases its unused checkpoint payload.
+
+## Qualification scope
+
+The reviewed external primitive candidate completed eight CUDA unit tests with
+no skips under an owned 120-second watchdog and a 70% allocator cap. This covers
+stock AdamW primitive equivalence, CPU state residency, malformed-state refusal
+and optimizer restore. Its unchanged optimizer math is distinct from the trainer
+integration: the integration additionally fixes canonical parameter-schema
+serialization to use JSON lists. CPU configuration tests preserve default/fused
+hashes and exercise cross-backend resume rejection using simulated device metadata.
+
+The integrated source at `b5d8299b768564c09c541e3dfc04b3d28eef2fd6` subsequently
+completed another owned primitive run: eight tests passed, zero skipped. The
+[portable tiny qualification receipt](evidence/optimizer/tiny-cuda-offload-v4-20260930.json)
+records all eight FP32/BF16 production-trainer process phases passing with no
+skips: uninterrupted two steps, durable step1 exit, fresh-process resume to
+step2, and reload/compare. All 33 production Python hashes and source HEAD
+matched at each process start/end. Every model and optimizer tensor element,
+scheduler, sampler, RNG and the whole training exposure matched exactly; resume
+attempt lineage and each checkpoint's own parent chain were checked separately.
+
+This is a four-row synthetic byte-tokenizer fixture with 30 parameters and
+28128 model tensor elements, not a pretrained fullweight result. Live persistent
+CPU optimizer state was 225144 bytes; comparison covered 56286 optimizer tensor
+elements. A zero tensor-element count for scalar-only fields means those fields
+were asserted equal, not omitted. Effective matmul precision was highest,
+TF32 was off, deterministic algorithms were enabled and the allocator cap
+was 70%. The earlier external candidate primitive receipt remains distinct.
+
+A separate tiny **LoRA rank4 with activation checkpointing** run completed all
+8/8 FP32/BF16 trainer phases with zero skips on the same frozen source. The
+[portable LoRA receipt](evidence/optimizer/tiny-cuda-offload-lora-checkpoint-v2-20260930.json)
+records exact model/optimizer/RNG/exposure and resume state comparisons,
+12 trainable parameters, 26 unchanged frozen states, and observed LoRA/head
+updates. Live CPU optimizer state was 15664 bytes. The integrated v4 primitive
+proof was reused; this was not another primitive test run. These are synthetic
+fixture results, with no native fullweight or acceptance claim.
+
+Actual pretrained full-weight step/save/exit/resume/reload and the
+captured-gradient stock optimizer oracle remain pending. Their fixed 16 GiB
+available-host-RAM admission requirement is currently unmet. Existing default
+backend OOM receipts remain failures. Primitive or CPU fixture checks establish
+no independent full-weight derivative parity, trained long context,
+representative quality, Jev advantage or issue closure.

docs/RELEASE_READINESS.md

@@ -1,5 +1,18 @@
 # Release readiness — intentionally incomplete
 
+## Explicit CPU-state optimizer: partial qualification (September 30, 2026)
+
+The opt-in backend at signed source `b5d8299b768564c09c541e3dfc04b3d28eef2fd6`
+passed the [tiny rank0 trainer qualification](evidence/optimizer/tiny-cuda-offload-v4-20260930.json)
+and a separate [tiny LoRA rank4/activation-checkpointing qualification](evidence/optimizer/tiny-cuda-offload-lora-checkpoint-v2-20260930.json).
+Each completed 8/8 FP32/BF16 uninterrupted/save/exit/fresh-resume/reload phases,
+with zero skips and exact state comparisons. All 33 source hashes and HEAD
+were unchanged at each phase start/end. The LoRA run reused the integrated v4
+8-pass primitive receipt; it did not rerun those tests. This is partial delivery
+using synthetic byte-tokenizer fixtures. Actual pretrained fullweight optimizer
+qualification and the captured-gradient stock oracle remain pending under their
+separate 16GiB host gate. Default OOM receipts and issues #3/#4 remain unchanged.
+
 ## Observed synthetic context decisions (September 30, 2026)
 
 The [development diagnostic](evidence/context/development-factorial96-20260930.md)

docs/acceptance-status.json

@@ -1,4 +1,21 @@
 {
+  "explicit_CPU_state_optimizer_partial_delivery_2026_09_30": {
+    "signed_source_commit": "b5d8299b768564c09c541e3dfc04b3d28eef2fd6",
+    "scope": "partial synthetic tiny-fixture optimizer/trainer delivery; no native/fullweight/context/quality acceptance",
+    "rank0_receipt": "docs/evidence/optimizer/tiny-cuda-offload-v4-20260930.json",
+    "lora_checkpoint_receipt": "docs/evidence/optimizer/tiny-cuda-offload-lora-checkpoint-v2-20260930.json",
+    "rank0_precision_phases_passed": 8,
+    "lora_precision_phases_passed": 8,
+    "phase_skips": 0,
+    "integrated_primitive_passed": 8,
+    "primitive_skips": 0,
+    "lora_primitive_scope": "reused from rank0 v4; not rerun",
+    "lora_trainable_parameters": 12,
+    "lora_unchanged_frozen_states": 26,
+    "remaining_gate": "actual pretrained fullweight step/save/exit/resume/reload and captured-gradient stock optimizer oracle pending; separate16GiB available-host gate",
+    "issue_3": "open",
+    "issue_4": "open"
+  },
   "context_development_diagnostic_2026_09_30": {
     "scope": "Observed synthetic short/medium development diagnostic; not acceptance, representative quality or Jev comparison",
     "receipt": "docs/evidence/context/development-factorial96-20260930.json",