[WIP] Add HyperFrames composition for Microsoft Learn video project

[Copilot]

• Inspect approved script, design plan, scaffold, and transcript anchors
• Fill chrome and author map-centric custom body scenes
• Assemble and run all required composition gates
• Review, security-scan, and finalize the verified project

Original prompt

Author the HyperFrames composition for the Microsoft Learn video project at learn/output/Highway-12-A-Lewis-And-Clark-Road-Trip-To-Montana/, following .github/copilot-instructions.md exactly (the build contract, anchoring convention, and gates all live there).

All inputs are already committed and FROZEN — do NOT change narration, palette, or the images:

• script.md — approved narration + beat plan + source-fidelity ledger (a balanced, NPS-sourced Lewis & Clark travelogue; keep the honest, reverent tone).
• design-plan.md — THE visual blueprint. Read it fully. The concept is map-centric.
• narration.wav (Dragon HD, ~372s) + transcript.json (971 words) + anchors.json (17 cues). Regenerate anchors.js with tools/word_anchors.py transcript.json --spec anchors.json -o anchors.js (it is gitignored, so you must regenerate it in the working tree).
• 14 generated images in assets/media/img/ (LFS): 3 portraits (jefferson/lewis/clark) + 11 landscapes.

THE VISUAL IDEA (do this, not static Ken-Burns photo scenes): ONE persistent animated route map is the stage. The route line DRAWS ITSELF leg by leg across WA→ID→MT, a marker rides the head, and as the drive reaches each waypoint the matching PHOTO CARD POPS UP beside it (scale 0.9→1 + fade + soft settle), holds under its narration, then recedes. History section pops the 3 portrait cards over a faint expedition map; the Sacagawea beat and all Native beats use LAND + attributed on-screen words, NEVER an invented face. The 1877 Nez Perce flight beat draws a second, mournful line the opposite way over Lolo Pass to the Big Hole. design-plan.md gives exact waypoint coordinates (viewBox 0 0 1920 1080), the image+cue+card-side per waypoint, and the chrome __FILL__ values.

Build steps: fill the 5 chrome __FILL__ scenes from design-plan.md; author the map body scenes on the block foundation (custom sub-comps, DOM/SVG/GSAP only — WebGPU fails the gate); anchor every card pop and label to its cue via window.__anchors (never invent offsets); insert body scenes into scenes.json in the body_slot; run tools/assemble_scenes.py, then tools/check_subcomps.py, tools/check_placeholders.py, tools/check_initial_state.py, tools/check_cue_anchors.py, then npx --yes hyperframes@0.7.77 lint and npx --yes hyperframes@0.7.77 check. ALL gates must pass. Do NOT render (Windows-only). Touch ONLY files under learn/output/Highway-12-A-Lewis-And-Clark-Road-Trip-To-Montana/. Open the PR with the gate results in the description.

[dahans-msft2]

Killing this run — restarting the build on Opus 5 via the UI model picker (MCP delegation can only use the default model).