[linson007]

Apple Silicon users can select --device mps for optional DFlash2 capture, preflight, training, resume, and final evaluation.

Depends on #499. Please merge the parent first. This upstream PR targets main, so it currently includes the parent training changes. The Mac-only follow-up is commit 2fe5fe39. After #499 merges, rebase that follow-up onto current main to isolate the final diff. This remains draft pending real-checkpoint qualification.

The Mac follow-up adds:

• MPS availability checks, teacher-cache release, synchronized timing, and preflight memory snapshots. Snapshots are explicitly distinguished from peak memory.
• MPS RNG checkpoint/restore and execution provenance covering Apple GPU, macOS, and explicit CPU fallback.
• Shared CPU BF16 conversion for export and MPS validation, with original FP32 optimizer parameter objects restored after successful or failed validation. This addresses a validation/export mismatch reproduced locally on MPS.
• Nine Mac-specific checks covering CLI selection, unavailable/required backend behavior, actual teacher generation and capture, selector and backbone optimization, exact resumed-versus-uninterrupted checkpoints, BF16 export, final evaluation, and failure recovery.
• A contributor testing guide and a manual workflow for an existing trusted Apple Silicon runner. The workflow requires registration on the default branch before dispatch; local smoke commands work directly from this PR.

Validation on an Apple M5 Pro with 48 GB unified memory, macOS 26.6, Python 3.12.14, PyTorch 2.13.0, and Transformers 5.15.1:

• All 50 optional training tests passed, including all nine Mac checks and pinned Z Lab reference parity, with explicit CPU fallback disabled.
• All 35 ordinary launcher tests passed.
• Without PyTorch, optional discovery passed 17 metadata/launcher checks with 33 expected skips.
• Ruff, formatting, compileall, architecture, workflow YAML/trigger checks, and whitespace checks passed.

The full ordinary/native suite was not rerun for the Mac follow-up. No real Qwen3.8-27B MPS training, production CUDA run, or trained-checkpoint native qualification is claimed.

Testing help wanted: contributors with sufficient unified memory for the full BF16 teacher can help measure real-checkpoint preflight, training memory/step time, resume, export, and native acceptance/throughput. Teacher weights alone are approximately 54 GB before working memory. The current command runs capture and training on the selected backend and does not provide cross-device cached-feature training.

The smoke test needs no model downloads after installing training/requirements.txt in a separate environment:

DFLASH_REQUIRE_MPS=1 PYTORCH_ENABLE_MPS_FALLBACK=0 python -m unittest dev.tests.training.test_mps -v

Please follow the guide's reporting checklist for hardware, versions, input revisions, recipe/split fingerprints, memory, timings, and errors.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}