Customize: Rain Station animated short

Make me a 30-second 2D animated storyboard previs short, "雨站" ("Rain Station"), with benchmark-level image quality. Don't install HyperFrames/Remotion; write all the code yourself (Python + ffmpeg).

Story (24fps, 2.39:1, 4K): an elevated train platform on a rainy night. She wears a beige trench coat and holds a clear umbrella, standing on the near platform; across the tracks under a streetlight stands him, in a long dark-gray overcoat and a scarf.
1. 0–6.4s Close-up of raindrops on the clear umbrella → pull back, focus shifts to the far side revealing him (over-the-shoulder shot)
2. 6.4–9.4s Medium close-up of him: head down → looks up toward frame left
3. 9.4–12.4s Frontal close-up of her: blinks → recognizes him; a horn sounds in the distance
4. 12.4–17s Wide shot: the headlights arrive first, a 5-car train passes; in the gap after the first car he's still there, by the second he's gone
5. 17–18.8s Close-up of her with tears in her eyes, bands of light from the train windows sweep across her face, her hair blown by the wind
6. 18.8–22.4s The taillights recede, the opposite platform is empty; his lamp flickers a few times and goes out
7. 22.4–28.6s Her back view, focus pulls from the empty platform back to her; footsteps behind her, a black umbrella covers her head, she turns and looks toward the camera; male voice: “好久不见。” (“Long time no see.”), with subtitles
8. 28.6–30s Title on black: “雨站 / RAIN STATION”

Approach:
- Artwork: use GPT Image 2.5 to draw the wide shot, him, her, and the train (lead car/middle car/rear car). Characters always on a pure green #00FF00 background and key them out yourself (asking for a transparent background with a reference image gives a fake checkerboard). Keep all poses of the same character in the same session, editing continuously with the base image as reference; when changing poses, composite only the changed region.
- Image quality: use Real-ESRGAN x4plus anime 6B (the safetensors version on HF; write RRDBNet yourself, run on MPS) to upscale all artwork 4x; use Depth Anything V2 to estimate depth for the wide shot and do per-pixel parallax.
- Rendering: write your own 2.5D compositor. Place each layer at its true depth, with a pinhole camera moving at real focal lengths; do circular bokeh, focus pulls, 3D rain streaks, train motion blur, volumetric light, bloom and film halation, fine grain, and color grading. Use mesh deformation on characters for breathing, wind-blown hair tips, and umbrella shake, switching key poses Japanese-style "on twos." 4K supersampling, cache frame by frame to disk, and parallelize across 3 processes by time range (16GB RAM).
- Sound: real Mixkit sound effects (rain, rain on the umbrella, train, horn, footsteps); write a score keyed to story beats and render it with the Steinway piano and string samples built into Logic/GarageBand; dialogue with a Kokoro Chinese male voice; per-shot sound distance, loudness -16 LUFS.
- At every step, render preview frames and tile them into a contact sheet to self-check; fix issues before outputting the film. Deliver a 4K master and a 1080p sharing version.

For best results, run it with Claude Opus 5.5.