D807 — A launch film on the Atlas engine leads the README

Decision. The README opens with a one-minute silent film instead of the Atlas tour: four teammates, each with their own coding agents, share one wiki; Monday one agent writes a fix down, Tuesday another agent on another client finds it, every day the others add to it and a broken link is refused. It is motion graphics written in film/, drawn by the Atlas's real graph engine (LivingScene offscreen, D691) and rendered frame by frame like the hero (D790): 1920x1080 at 60 fps to MP4 for social posts, 1120 px at 30 fps to WebP for the README. film/publish.sh force-pushes the WebP as the single commit of the orphan branch readme-film. The Atlas tour stays, under "What you get".

Why. The tour shows the Atlas, not what Cartographer is for: nothing on the page showed several people and several clients sharing what their agents learn, which is the whole pitch. A recording of real sessions was considered first, but the story needs two people, two machines and a day between them, which a screen capture cannot show in a minute; staged motion graphics can, and drawing the graph with the real engine keeps it honest about what the Atlas looks like. The film costs a manual re-render and publish when the story or the look changes, and it is reviewed by nobody in a PR, as the hero is.

Alternatives rejected. - A screen recording of real Claude Code and Codex sessions. It cannot show a team or a day passing, and the agents' output is not reproducible between takes. - The graph drawn by the brand kit's studio prototype. Cleaner, but it would show a graph the product does not draw. - The film on readme-assets beside the hero. publish-hero.sh force-pushes a tree holding only the hero, so every release that changes web/ would delete the film. - The WebP committed to main. ~14 MB per version in every clone, the cost D790 removed for the hero. - A dark-theme variant. The paper look is the film's identity and reads as a card on a dark page; a second render would double the work and say nothing more. - A soundtrack. A synthesised one was built and dropped: the README plays the WebP muted anyway, and the social cut does not need it. - 60 fps in the WebP. Browsers do not honour animated-image frame delays much under 20 ms. - Recording the film in CI on every release, like the hero. Worth doing, but the story changes by hand and rarely; deferred to a follow-up.

Consequences. The film lives outside web/, so editing it does not touch the Atlas bundle's provenance hash; it imports the engine from web/src and calls one private method of scene.ts (animateTo, commented where it is used), so a refactor of the camera there can break the film's render without failing a test. The README depends on a second branch outside main's history, readme-film.