Replay a demo script on a virtual display and write a silent MP4 with step timings

Problem

There is no way to record an agent's demo headlessly and reproducibly: the existing screen-recording skill assumes
OBS on a person's desktop.

Change

• The container image: virtual display, Chromium, Playwright, ffmpeg.
• The npm package flow-reel with the demo() / step() API (story decision 010) and the CLI: on the host,
  flow-reel record demo.ts starts the image with the project mounted; inside, it replays the script and records.
• The runner's stabilisation from story decision 009: fixed viewport and device scale, prefers-reduced-motion, fixed
  clock and seed, network-idle wait before each step.
• Recording with Xvfb + ffmpeg into a 4K near-lossless master at 30 fps, plus a silent 1080p demo.mp4 (decision 001).
• A wrapped page that eases the mouse to each target and types visibly, so later polish can draw the cursor
  (story decision 012, this ticket's decision 002).
• transcript.json with video and each step's index, startMs, actionEndMs, endMs (story decision 011), and
  events.json with the runner's moves, clicks, typing and scrolls (story decision 013).
• Reaching an app on the host's localhost on Linux and Docker Desktop (decision 003); the image tagged with the npm
  version, built locally when it can't be pulled (decision 004).

Out of scope

Narration, captions, audio (GH-3 to GH-5); drawing the cursor, zoom and SFX (later tickets); publishing images
(the release ticket).

Done when

• flow-reel record on a sample app writes a 3840×2160 30 fps master, a silent 1920×1080 demo.mp4, and a
  transcript.json and events.json that validate against version 1.
• Two runs of the same script have the same steps, every start within ±150 ms.
• A step's logged startMs matches the frame where it visibly starts, within ±2 frames.

Decisions

This ticket's decisions 001-004 (recorder, wrapped page, reaching the host, image).