docs: split m05 external fastmtp parity

Author: apstenku123

bench/parity/m05_fastmtp.json

@@ -116,7 +116,7 @@
   "gb10_mtp_parity_claim": false,
   "gb10_parity_claim": false,
   "issue": {
-    "id": "cppmega-mlx-t8f.5",
+    "id": "cppmega-mlx-hjfn",
     "title": "M0.5: fail-closed FastMTP parity scaffold"
   },
   "local_only": true,
@@ -128,7 +128,7 @@
   "m4_vs_gb10_mtp_parity_claim": false,
   "m4_vs_gb10_parity_claim": false,
   "metadata": {
-    "git_commit": "2728cfe",
+    "git_commit": "7315eb6",
     "metadata_preflight_only": true,
     "scaffold_only": true,
     "script": "scripts/m05_mtp_parity_manifest.py"

cppmega_mlx/training/parity.py

@@ -125,7 +125,7 @@
     "valid_not_evaluated": "refused_not_evaluated",
 }
 M05_MTP_PARITY_RECEIPT_SCOPE = "local_cuda_mlx_m05_fastmtp"
-M05_MTP_PARITY_ISSUE_ID = "cppmega-mlx-t8f.5"
+M05_MTP_PARITY_ISSUE_ID = "cppmega-mlx-hjfn"
 M05_MTP_PARITY_OUTPUT = "bench/parity/m05_fastmtp.json"
 M05_MTP_PARITY_PROFILE = "local_gb10_quarter"
 M05_MTP_DEPTH = 2

docs/mlx_port_master_plan.md

@@ -237,7 +237,7 @@ Before fanning out across all 10 streams, prove the smallest viable path on the
 - M0.2 — Model factory entry for local_gb10_quarter in cppmega_mlx/recipes/model_factory.py. With M0.1 closed, the default profile is unblocked; acceptance remains the profile contract plus build → forward closure on shape (B=1, T=512) returning finite logits with config schema validation rejecting invalid combos.
 - M0.3 — **External reference closure, not a Mac-local blocker**. Random-init seed-matched forward parity remains useful for comparing MLX against a CUDA/GB10 reference, but it does not gate local MLX/Metal M0.4 training progress. The fail-closed scaffold constructs the same local_gb10_quarter config and fixed B=1,T=512,seed=3003 input batch (tokens_sha256=c645ca4053e5206dcbe58c13aa26f4a9e56c5aa2aee90a4d4778bbc9d9c33549). No real CUDA logits artifact exists at bench/parity/cuda/m03_local_gb10_quarter_seed3003_logits.json, so scripts/m03_forward_parity_manifest.py must refuse the default missing artifact and keep m0_3_closed=false. A metadata-valid CUDA artifact only reaches artifact_preflight_status=valid_not_evaluated; it is not external reference acceptance until a separate numerical harness compares the full logits tensor (shape=[1,512,65536], numel=33554432) against MLX and records pass/fail parity. The non-blocking bd follow-up is cppmega-mlx-uwhj; the old Mac-local blocker task cppmega-mlx-t8f.3 is superseded by that external-reference task, not closed as completed parity evidence.
 - M0.4 — One training step in bf16 (loss + backward + optimizer.update). No NaNs. Loss decrease over 100 steps on the target local parquet sample. **Data source**: local ignored data/parquet_samples/gb10/clang_semantic_4k_v10/val_00000.parquet (~492 MB total of GB10 validation shards when present locally; not committed to git). Full local_gb10_quarter acceptance is captured in bench/baselines/m04_train_step.json with --grad-checkpoint on the target parquet: 100/100 steps, loss 11.2004585 -> 6.9402304, finite loss/grad path, AdamW with FP32 master moments, mean step time 10.72670 s, 16.28 target tokens/s, and peak process memory 23,917,647,383 bytes. No GB10 scp or full-corpus prep needed for M0.
-- M0.5 — MTP K=2 head wired with β=0.6 / λ=0.3; per-depth losses tracked. MTP-disabled inference path also returns sane logits. Current local coverage is bounded to MLX training-side loss semantics plus fail-closed manifest hardening, including nested CUDA/MLX overclaim rejection. External CUDA/GB10 FastMTP loss+grad-norm closure remains open until a fixed-seed artifact exists and an MLX-vs-CUDA numerical harness evaluates it; that external reference closure does not block Mac-local M0.4 training acceptance.
+- M0.5 — MTP K=2 head wired with β=0.6 / λ=0.3; per-depth losses tracked. MTP-disabled inference path also returns sane logits. Current local coverage is bounded to MLX training-side loss semantics plus fail-closed manifest hardening, including nested CUDA/MLX overclaim rejection. External CUDA/GB10 FastMTP loss+grad-norm closure remains open until a fixed-seed artifact exists and an MLX-vs-CUDA numerical harness evaluates it; that external reference closure does not block Mac-local M0.4 training acceptance. The non-blocking bd follow-up is cppmega-mlx-hjfn; the old mixed-scope blocker task cppmega-mlx-t8f.5 is superseded by that external-reference task, not closed as completed FastMTP parity evidence.
 - M0.6 — Memory math validated on the actual dev box (Mac Studio M4 Max 128 GB): peak unified-memory < 75% of installed RAM with --grad-checkpoint and AdamW.
 - M0.7 — Resumable training: save mid-run, kill process, reload, identical loss continues for ≥100 steps with RNG state and dataloader cursor preserved. Current receipts are TinyLM/HybridTinyLM-scoped only; full M0.7 remains fail-closed until local_gb10_quarter resumes on the target parquet lane after the M0.4 full-model training gate is proven.
 

docs/parity_anchors.md

@@ -441,10 +441,13 @@ launchers.
   production cadence scheduling, sequence-parallel gather, or CUDA/Hopper/GB10
   kernel parity. Treat Hopper fused CE as source-only evidence until a separate
   hardware-gated receipt exists.
-- Fail-closed M0.5 receipt: no checked-in or locally observed fixed-seed
+- Fail-closed M0.5 receipt: the active non-blocking external-reference bd task
+  is cppmega-mlx-hjfn; the older mixed-scope blocker cppmega-mlx-t8f.5 is
+  superseded by that task, not closed as completed FastMTP parity evidence. No
+  checked-in or locally observed fixed-seed
   CUDA/GB10 artifact currently records the required FastMTP loss values plus
   grad norm for cppmega/megatron/fastmtp_layer.py and
-  cppmega/megatron/mtp_native_hopper_ce.py. M0.5 remains open until that
+  cppmega/megatron/mtp_native_hopper_ce.py. External M0.5 remains open until that
   artifact exists and a numerical MLX-vs-CUDA comparison is evaluated; the
   local tests above are contract coverage, not closure evidence.
 - Hardening receipt: the M0.5 manifest builder/validator now rejects nested

tests/test_m05_mtp_parity_manifest_script.py

@@ -26,6 +26,7 @@
 ROOT = Path(__file__).resolve().parents[1]
 SCRIPT = ROOT / "scripts" / "m05_mtp_parity_manifest.py"
 PYTHON = ROOT / ".venv" / "bin" / "python"
+EXTERNAL_M05_FASTMTP_TASK_ID = "cppmega-mlx-hjfn"
 
 
 def run_script(*args: str, timeout: int = 30) -> subprocess.CompletedProcess[str]:
@@ -87,6 +88,10 @@ def write_json(path: Path, payload: Any) -> None:
     path.write_text(json.dumps(payload) + "\n", encoding="utf-8")
 
 
+def test_m05_manifest_tracks_current_external_reference_task() -> None:
+    assert M05_MTP_PARITY_ISSUE_ID == EXTERNAL_M05_FASTMTP_TASK_ID
+
+
 def assert_m05_fail_closed_invariants(payload: dict[str, Any]) -> None:
     validate_m05_mtp_parity_manifest_dict(payload)
     assert payload["issue"]["id"] == M05_MTP_PARITY_ISSUE_ID