Fix canonical FP32 gradients and BF16 branch graph lifetime

MANIFEST.in

@@ -7,6 +7,8 @@ recursive-include scripts *.py
 recursive-include tests *.py *.txt *.json
 recursive-include integrations *.py *.json *.md *.ts *.tsx
 recursive-include clients *.ts *.json *.mjs
+prune clients/typescript/node_modules
+prune clients/typescript/dist
 recursive-include evals *.json
 global-exclude *.pyc
 global-exclude *.pt *.pth *.safetensors *.onnx *.ckpt

docs/RELEASE_READINESS.md

@@ -1,5 +1,23 @@
 # Release readiness — intentionally incomplete
 
+## Full-weight mechanics and graph lifetime supplement (September 29, 2026)
+
+The [graph lifetime repair and measured rank-0 limits](evidence/native/graph-lifetime-fullweight-limits-20260929.md)
+permit the same-init BF16 first step and reload, with a passing CUDA ownership
+regression and failing old-source control. Rank-0 FP32/BF16 resume backward still
+exhausts the unchanged allocator cap on this configuration. These results do not
+close #3/#4 or alter the source identity of earlier numerical/rank-8 receipts.
+
+## FP32 candidate numerical readback (September 29, 2026)
+
+The [canonical FP32 repair](evidence/native/short-fp32-canonical-v10-20260929.md)
+passes the short independent HF adapter oracle, four generated cached/batched/full
+cases, 8191/8192 state-prefix boundaries, and a 10520-token state numerical gate.
+This is candidate source evidence; the exact-limit FP32 training resource probe exhausted the fixed allocator cap
+during forward, leaving backward unverified. Delivery readback remains pending. Issues #3/#4 remain open; representative quality and Jev benefit
+are unknown. Overwatch remains archived by user instruction. Older readbacks
+below are historical and do not override these scoped receipts.
+
 ## Current delivery readback (September 28, 2026)
 
 Layev main is `a6926f7b4585e840b36f1d9157c3af3907210ef5`, after protected

docs/acceptance-status.json

@@ -1,4 +1,23 @@
 {
+  "fullweight_graph_lifetime_supplement_2026_09_29": {
+    "receipt": "docs/evidence/native/graph-lifetime-fullweight-limits-20260929.json",
+    "graph_lifetime": "CUDA positive passes; proper old-source negative fails; 14 BF16 tests pass",
+    "rank0_fp32": "first step and reload pass; resume backward CUDA OOM; allocator-option retry fails",
+    "rank0_bf16": "same-init graph-repaired first step and reload pass; resume short-attention backward CUDA OOM",
+    "scope": "Generated short mechanics; measured configuration only; no complete full-weight matrix or quality claim",
+    "issue_3": "open",
+    "issue_4": "open"
+  },
+  "short_fp32_canonical_v10_candidate_2026_09_29": {
+    "scope": "candidate source short FP32 numerical repair only; no issue closure",
+    "receipt": "docs/evidence/native/short-fp32-canonical-v10-20260929.json",
+    "hf_adapter_gradients_passed": 96,
+    "generated_seam_cases_passed": 4,
+    "gradients_per_comparison": 100,
+    "long_validation": "fresh 8191/8192 prefix boundaries and 10520-token state numerical parity passed; exact-limit FP32 training forward OOM at unchanged 70% cap; backward unverified",
+    "issue_3": "open",
+    "issue_4": "open"
+  },
   "schema_version": 1,
   "kind": "workstream-evidence-ledger-not-original-feature-matrix",
   "base_revision": "b4cb9df97be61729ebd68b8e0f5ee9bb2d193c73",

docs/evidence/native/FP32_REPRODUCTION.md

@@ -0,0 +1,23 @@
+# Reproduce the scoped FP32 numerical evidence
+
+Use the CUDA model environment with PyTorch 2.10.0+cu128 and Transformers 4.57.1.
+Supply the locally retained trained native LoRA checkpoint; weights are not
+included in this repository. Both scripts record runtime source identity and
+refuse overwriting result artifacts. Output paths below must be fresh.
+
+```shell
+python scripts/verify_native_fp32_gradients.py --checkpoint CHECKPOINT.pt --expected-checkpoint-sha256 99f797ffd0521adbee964d130d73eb381c9b6bd3b7896ce61b5007886d19ccf9 --out fp32-hf.json
+python scripts/verify_native_fp32_seams.py --checkpoint CHECKPOINT.pt --expected-checkpoint-sha256 99f797ffd0521adbee964d130d73eb381c9b6bd3b7896ce61b5007886d19ccf9 --out-dir fp32-seams
+python scripts/verify_native_fp32_seams.py --checkpoint CHECKPOINT.pt --expected-checkpoint-sha256 99f797ffd0521adbee964d130d73eb381c9b6bd3b7896ce61b5007886d19ccf9 --out-dir fp32-long --cases boundary_8191 boundary_8192 long_10520
+```
+
+The first command compares hidden states and all Q/V LoRA parameter gradients
+against independent eager HF parameters. The second runs four exact-token
+cached/batched/full numerical cases. The third constructs the same generated
+8191/8192 boundary and 10520-token state requests and applies the same loss and
+three-mode backward protocol. It requires substantial GPU time. The 70% CUDA allocator cap and disabled
+TF32 are fixed. HF hidden tolerance is 1e-4, logits/probabilities 1e-5, and
+trainable gradients 2e-5 (combined absolute/relative). Generated requests
+establish no representative quality, long-context or Jev benefit. The seams
+protocol retains its explicit 16-GiB cache-policy ceiling; this is a protocol
+planning ceiling, not allocator budget growth or a total-memory claim.