M8: 744B QLoRA on 64 GB — evidence

Evidence for the M8 milestone: QLoRA fine-tuning of GLM-5.2 (744B params,
384 GB int4) on a single 64 GB M4 Max, produced by scripts/m7_smoke.sh and
scripts/m8_overfit.sh at this tag.

Attached:

• m7_smoke.log — M7 smoke test: build → doctor → 1+10 training steps →
  adapter loaded in inference (loss finite throughout, peak RSS 34.8/44 GB)
• m8_overfit.log — the full M8 run: 100 steps / 10 epochs, loss 5.05 → 0.37,
  27.8 h, per-microbatch heartbeats and [mem] telemetry
• m8_eval_generations.txt — greedy base-vs-adapter generations: train prompts,
  the held-out prompt, and base behavior restored with the adapter off

Upstream PR: JustVugg/colibri#626