docs: record nonblocking m03 external parity task

Author: apstenku123

docs/mlx_port_master_plan.md

@@ -235,7 +235,7 @@ Before fanning out across all 10 streams, prove the smallest viable path on the
 **Gate set** (Mac-local gates must be green before scaling effort; external reference closure is tracked fail-closed but does not block Mac-local M0.4 work):
 - M0.1 — **CLOSED**. Tokenizer: vendor the deployed GB10 tokenizer artifact with vocab=65536 and the reserved ID contract (id 2=BOS, 3=EOT/EOS, 4=FIM_PREFIX, 5=FIM_MIDDLE, 6=FIM_SUFFIX, 7=CODE_START, 45=FIM_INSTRUCTION, 46=SPACE, 47=NL). The wrapper normalizes whitespace runs at encode ([\r\n]+-><NL>, [ \t]+-><SPACE>) and decode is plain token concat with sentinel substitution; this fixes the BPE-split decode bug (e.g., sum->s,u,m->s u m) and gives byte-exact round-trip for inputs without multi-char whitespace runs. The explicit-token approach matches the CUDA-side nanochat/cpp_tokenizer.py decode behavior. Do not reopen this gate over the deployed HF decoder=null artifact's non-reversible decode(encode(text)) behavior.
 - 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.
+- 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.
 - M0.6 — 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.