docs(hero): 4x sped-up README hero GIF

[juacker]

Replaces docs/clai-hero.gif (the README hero) with a 4× sped-up re-encode of the 115 s landing capture. No README change: the path is unchanged.

	before	after
duration	16.0 s	28.75 s
size	2.69 MB	3.34 MB
resolution	860×484	640×340
fps	10	12
loop	infinite	infinite (NETSCAPE2.0, loop_count=0)

Why the filter chain looks like this

The source capture is variable frame rate — measured PTS gaps range from 0.033 s to 0.446 s (~20.7 fps average). setpts=PTS/4 alone works on presentation timestamps and does not normalise variable timing, so the chain resamples to CFR first, then scales PTS:

ffmpeg -y -i video-landing.mp4 \
  -vf "fps=30,scale=640:-2:flags=lanczos,setpts=PTS/4,fps=12,split[s0][s1];\
[s0]palettegen=stats_mode=full[p];[s1][p]paletteuse=dither=sierra2_4a" \
  -loop 0 docs/clai-hero.gif

Deviation: 640 px wide at 12 fps, not 720 px at 15 fps

720/15 came out at 4.92 MB — under the 5 MB ceiling but far from the ≤3 MB ideal and ~83% heavier than the asset it replaces. Measured fallback ladder (width → fps → dither, as specified):

variant	size
720 × 15, sierra2_4a	4.92 MB
720 × 15, bayer_scale=5	4.73 MB
720 × 15, 128-colour palette	4.36 MB
720 × 12, sierra2_4a	4.25 MB
640 × 15, sierra2_4a	3.89 MB
640 × 12, sierra2_4a	3.34 MB ← chosen
640 × 10, sierra2_4a	3.29 MB

Bayer dithering barely helped on this content (dense UI text), so it was not used. 640×12 is the smallest variant that stays recognisable: dropping to 10 fps saves only 0.05 MB and makes the already-4×-fast motion visibly choppy, so 12 fps is the stopping point. The GIF is rendered at width="100%" in the README, so the narrower intrinsic width costs no layout change.

Verification

• 345 frames, 28.75 s, 640×340 (even dimensions), infinite loop confirmed by parsing the NETSCAPE2.0 application extension (loop_count=0) and the 0x3B trailer.
• Temporal mapping is 4×: source t=80 s matches GIF t=20 s at mean |diff| 0.52/255 (palette dither noise only).
• Full decode is clean; the GIF covers the whole flow from the first to the last frame of the source.