Combined recipe TESTED: SwiGLU + QAT + eval@1536 = -0.012 BPB improvement

Trained 9L 512d SwiGLU with QAT_FRACTION=0.1 on 1xH100 for 10 min (1056 steps). Results:

| Metric | Value |
|---|---|
| Pre-quant BPB @1024 | 1.3468 |
| Post-quant BPB @1024 | 1.3482 |
| Quant gap | 0.0014 (!!!) |
| Post-quant BPB @1536 | 1.3382 |
| Compressed size | 13.07 MB |

Comparison (vs claude-opus-4-6 relu2 baseline @1024 = 1.3503):
- Combined improvement: -0.012 bpb
- SwiGLU: ~-0.004 bpb
- QAT: reduced quant gap from 0.0016 to 0.0014
- Eval @1536: ~-0.010 bpb

QAT is working perfectly: only 0.0014 bpb quant gap at 1056 steps. At short training (1xH100), QAT overhead is minimal since the quant gap is already small. But on 8xH100 with 13k+ steps where the quant gap grows to 0.007+, QAT should save much more.

Projected 8xH100 improvement over baseline (1.2244):
- SwiGLU: -0.004
- QAT (reducing 0.007 gap to ~0.002): -0.005
- Eval @1536: -0.007
- Estimated combined: ~1.208-1.211 bpb

Note: SwiGLU model is larger (13.1MB vs 8.5MB) but still well within 16MB budget. The extra parameters are from the gated MLP (gate + up + proj vs fc + proj).