M4 Pro 24GB: fold Turbo LoRA (PR #14) benchmark — 124f is the sweet spot on both 480P and 768P

# M4 Pro 24GB: fold Turbo LoRA (PR #14) 4-spec benchmark — 124f is the sweet spot on both 480P and 768P

Machine: Mac mini M4 Pro, 24GB unified memory, macOS 26.6.2, `--ssd-streaming`.
Test: `MiniMax-H3-turbo` (fold Turbo v4), FL2VA first/last frames, seed 42, 6 steps,
ET5T shot 1 (strong-motion drift, gravel dust). Wall-clock = process birth → `h3: wrote mp4`.

| Resolution | Frames | Steps | Wall-clock |
|---|---|---|---|
| 832×480 | 124 | 6 | 19.4 min (01:12→01:31) |
| 832×480 | 243 | 6 | 46.0 min (07:31→08:17) |
| 1344×768 | 124 | 6 | 75.5 min (08:47→10:03) |
| 1344×768 | 243 | 6 | 206.9 min (~3.4 h) (10:50→14:17) |

Key observations:

1. **124 frames is the sweet spot for BOTH 480P and 768P.** Per-frame
cost and wall-clock stay predictable at 124f; jumping to 243f is
clearly superlinear on this memory budget. If you need ~10s, plan for
a near-linear cost jump in time (and a healthy swap peak).

2. **Single-step linear extrapolation fails for long sequences.** Naively
scaling 768P/124f (~11 min/step) by latent ×1.9 predicts ~21 min/step
for 768P/243f; measured (incl. load) is ~34 min/step. ETA for folded
long shots should not be extrapolated linearly — wait for step 1 and
re-baseline.

3. **768P fold 6 steps runs fine on 24GB** — no OOM, no major swap
(peak swap ~11.5 GB). Upscaling from 480P+super-res is still cheaper
than native 768P/243f (~3.4 h) on this box, but native 768P is viable.

Optional quality note: on this strong-motion clip, Laplacian variance
was lower for fold (309.4) than 4-bit mlx-serve (604.0), yet human
subjective judgement favored fold's natural sharpness — the mlx output
reads as over-sharpened artificial halos. So Laplacian alone is not a
quality proxy; "natural" vs "over-sharp" must be judged by eye.