feat(script): cinematography agents + asset library + hook composer

[cuio]

Summary

Three things in one PR — the planner stops being template-only and starts directing cinematography, scenes get a much richer visual kit to draw on, and a cheap second-pass critic protects the cold-open.

Stacked on PR #7 — branch off it because slice 8 needs the atmosphere/transition registries it adds.

Asset library

• 4 new atmosphere presets (registry: 5 → 9):
  • radial-pulse — concentric ring pulses from canvas centre. Magnetises the eye for centred hero elements (hook-statreveal).
  • cosmic-dust — sparse twinkling stars + two drifting halos. Cosmic scale feel for hooks/outros.
  • geometric-grid — drifting + pulsing isometric grid lines. Engineering aesthetic for chart-scene and comparison.
  • flow-lines — wavy horizontal SVG paths drifting across the canvas. Audio-waveform vibe for quote scenes and time-series charts.
• Icon library — new packages/core/src/script/icons/ with 20 hand-drawn 24×24 stroke SVGs (lock, network, bolt, dollar, globe, target, shield, etc). Pure geometric primitives, no third-party paths. Renders at any size, picks up token colours via currentColor.
• concept-callout items can now carry an icon that replaces the number badge: { text, icon } per item, mixed with plain strings allowed. Backwards-compatible.

Slice 8 — planner cinematography

• New "Cinematography" section in the playbook teaches the AI when to pick each atmosphere preset, which transition fits which moment, and which icon belongs on which list item.
• Planner tool schema gains optional background and transition fields per scene, both as enums of registered ids. Unknown values fall back to defaults so a hallucination can't break assembly.
• Schema validation + one-shot retry: after the first planner pass, every scene is checked for missing required props or chart-scene without chart.type / chart.props matching the chart's required fields. Issues are sent back as a corrective system prompt for ONE retry — turns ~90% first-pass success into ~99% at the cost of one extra round-trip on the failing case. Lingering issues land in meta.warnings so the UI can surface them.

Slice 9 — hook composer pass

• New improveHook(script, opts) that critiques s01 against the hook quality checklist and either keeps it or proposes a swap with a stronger sentence later in the script. Cheap (~1k input + 200 output tokens, ~$0.01/video on Sonnet 4.6).
• Verbatim fidelity skips by contract — strict mode never swaps. Refine and split-merge run the critic.
• Wired into POST /projects/:id/script/plan as an optional second pass, default on. Body flag improveHook=false skips. Decision and reasoning surface in meta.warnings.
• Failures are non-fatal: a failed critic call returns the original script with a reasoning string instead of bubbling.

Test plan

• All four new atmosphere presets render valid HTML with class + keyframes (verified via direct invoke)
• Icon registry has 20 icons; renderIcon produces inline SVG with stroke colour applied
• concept-callout template emits SVG inside .co-badge when items pass {text, icon}, two-digit number when items pass plain string
• Reassembled my-first-video with explicit props.background overrides on different scenes — geometric-grid renders crisp diagonal accent lines with vignette; cosmic-dust shows dual halos drifting
• Lint, format check, typecheck all pass
• Schema validator + retry path — type-safe; manual end-to-end requires an Anthropic key call to exercise

Notes on cost

• Planning a typical 60s video: ~5k input + 4k output tokens on Sonnet 4.6 — ~$0.07
• Hook composer pass: ~1k input + 200 output — ~$0.01
• Schema-correction retry (only when first pass fails validation): another ~5k+4k tokens — adds ~$0.07 on the ~10% of plans that fail validation
• Net: ~$0.08–$0.15 per planned video, on the high end with retry

Stack

• This PR depends on #7 (the atmosphere + transition registries it builds on)
• It composes well with #6 (captions UX) and #8 (scale-to-fit + DESIGN.md overlay) — they're all independent.