Investigate and surface status of PR #5 (M4 training script)

[Copilot]

PR #5 ("Update training script for local development on M4 MacBook Air") was opened with only a plan commit — zero code changes merged. All 11 checklist items remain unimplemented.

Findings

PR #5 state:

• Branch: copilot/implement-smoke-test-upgrades
• Commits: 1 ("Initial plan") — no files changed
• All checklist items unchecked

Outstanding work in PR #5

• Section 0 — M4/MPS device detection, DDP bypass, MPS-safe autocast, drop fused=True on AdamW
• Section 1 — SparseMoEMLP (top-k gating, aux load-balance loss) + Block integration
• Section 2 — CurriculumTokenLoader (easy-to-hard shard ordering, CURRICULUM_FRAC)
• Section 3 — ValueEmbedding (token identity reinjection into late-layer V, VALUE_EMBED_LAYERS)
• Section 4 — IBPhaseMonitor (gradient-norm ratio fitting/compression phase detector)
• Section 5 — check_router_health watchdog (MoE load-collapse detection)
• Section 6 — fake_quant_ternary + pack_ternary_np/unpack_ternary_np (base-3 packing, ~1.58 bits/weight)
• Section 7 — artifact_size_guard (hard budget check + 500KB warning threshold)
• Optimizer split updates for new param groups (router, value_embedding.*)
• Smoke-test summary line at startup

No changes have been committed to the repository as part of this investigation.

[kailean]

continue