What does this PR do?

Adds concise, agent-readable model verification cards under examples/model_verification_cards/ for four variants:

• Qwen/Qwen3-8B
• Qwen/Qwen3-30B-A3B
• moonshotai/Moonlight-16B-A3B
• nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16

It also adds the repo-local create-model-verification-card skill, its deterministic validator, and a helper for exact-length, byte-stable HF inference verification.

The cards distinguish bounded convergence evidence from benchmark-only performance evidence, pin immutable Hugging Face and dataset revisions, preserve recipe-owned GBS/MBS/LR contracts, require saved post-setup runtime configuration, and keep private executor wiring out of tracked cards and PR text.

Card contract

• Keep all twelve core items; include pretrain_performance only when the exact variant has a canonical public performance recipe.
• Start untuned cards with the required sanity-check disclaimer; scope tuned claims to exact pretrain_performance.<hardware> leaves.
• Use portable public launchers and short logical work/... paths.
• Pin immutable model, tokenizer, and dataset revisions plus split, seeds, and packing alignment.
• Keep recipe GBS/MBS/LR unchanged in verification commands.
• Record initial/final loss plus final-ten mean step time and model TFLOP/s/GPU for every verified training item.
• Nest hardware-dependent training and SFT export evidence under public accelerator keys (for example, pretrain.H100), with resume/export dependencies resolved only within the same hardware target.
• Require exact-length, byte-stable inference output from two executions.
• Verify manual forward correlation with an exact next-token match and cosine similarity at least 0.99; maximum and mean absolute logit differences remain report-only diagnostics.
• Resume directly from the uninterrupted bounded-pretrain step-50 checkpoint to step 100 in a distinct output root.
• Treat submission or a short smoke as evidence-in-progress, not verification.

Current verification status

The branch remains a draft while the final clean-contract resume/export audits finish. Current head: fbd1ad6a4.

The three long-context items are verified from one coherent clean-run evidence cohort:

Each completed exactly 20 optimizer steps with finite losses, zero skipped/NaN iterations, a persisted runtime config, and independently reconstructed packed-sample accounting.

GPU export config correction

The final commit fixes GPU export for checkpoints whose training-time config differs from the original Hugging Face config. Export retains the existing state-backed reference wrapper and shard map, but replaces its config with the checkpoint-derived Hugging Face config before saving. This preserves strict/distributed streaming behavior while making serialized config shapes agree with the checkpoint tensors.

The focused GPU-export tests pass 10/10 on both macOS and the fixed Linux container. Repository-wide pre-commit passes in that Linux container. The current public docs, secrets, and DCO checks are green.

Revision hardening and validation

• Hugging Face model and tokenizer revisions are persisted by the affected recipes instead of relying on mutable defaults.
• The revision-hardening patch passed 184 targeted tests and repository-wide pre-commit.
• The current long-context/card update passed all three card validators, tokenizer unit tests, Ruff checks, the CI-equivalent secrets scan, and repository-wide pre-commit.
• Two public immutable revision literals triggered the generic high-entropy detector; narrow inline allowlist annotations scope those known false positives to the exact literals.
• git diff --check passes, and independent evidence reviews reproduced every promoted metric and supervised-token count.

Remaining work

The final bounded-pretrain direct resumes and corrected Moonlight export are queued from frozen, independently reviewed launchers. Execution state alone does not promote an item: each remains unverified until its exact optimizer-step set, finite-loss/no-skip gates, saved runtime config, checkpoint or adapter artifacts, packing/supervised-token evidence, resume sentinels, strict load, deterministic inference, and frozen evidence audit all pass. The cards, PR status, and Linear issues will be promoted only after those gates complete.