Expand adaptive learning experience

.env.example

@@ -1,2 +1,11 @@
 VITE_INSFORGE_URL=https://your-project.region.insforge.app
 VITE_INSFORGE_ANON_KEY=your-anon-key
+
+# Server-only. Put the real value in .dev.vars locally and use
+# `npx wrangler secret put NVIDIA_API_KEY` for a deployed Worker.
+NVIDIA_API_KEY=nvapi-replace-with-your-key
+
+# Server-only Codex lesson generation. Put the real value in .dev.vars locally
+# and use `npx wrangler secret put OPENAI_API_KEY` for a deployed Worker.
+OPENAI_API_KEY=sk-replace-with-your-key
+OPENAI_LESSON_MODEL=gpt-5.6-sol

DEMO_SCRIPT.md

@@ -2,6 +2,16 @@
 
 The video must be public on YouTube, under three minutes, and include spoken coverage of how GPT‑5.6 and Codex were used.
 
+## Before recording — fast AI smoke path
+
+1. Keep the Codex desktop app open and signed in, then run `npm run dev` and open the printed local URL.
+2. Sign in to Perception Learning.
+3. Generate a lesson with topic **Why gas pressure rises when volume decreases** and goal **Explain the pressure change using evidence from a controlled simulation**.
+4. Open the generated ideal-gas simulation and change volume to show that only a reviewed deterministic runtime executes.
+5. In the lesson notes, type **Why does pressure increase when volume decreases?**, choose **Generate AI answer**, and show the answer appended on a new line. Generate study or multiple-choice questions from the same lesson as a second quick proof.
+
+Use the deployed URL for the normal hosted demo. Use `npm run dev` when you specifically want to demonstrate lesson/simulation generation through the signed-in Codex GPT‑5.6 bridge. NVIDIA question and note-answer generation works in both environments.
+
 ## 0:00–0:20 — The problem
 
 **Show:** Library hero and OpenStax shelf.
@@ -24,7 +34,7 @@ The video must be public on YouTube, under three minutes, and include spoken cov
 
 **Show:** Class → Start assignment → commit a Mechanical Energy forecast → run the lab → change release height → explain the conserved quantity.
 
-**Say:** “The model is interactive, repeatable, keyboard- and touch-friendly. AI can propose a model, but only validated deterministic code produces the evidence. The learner’s forecast, confidence, parameter changes, and explanation form the proof trail.”
+**Say:** “The model is interactive, repeatable, keyboard- and touch-friendly. AI can select and frame a reviewed runtime, but only allow-listed deterministic code produces the evidence—no generated JavaScript reaches the browser. The learner’s forecast, confidence, parameter changes, and explanation form the proof record.”
 
 ## 1:48–2:20 — The feedback loop that did not exist before
 
@@ -42,6 +52,6 @@ The video must be public on YouTube, under three minutes, and include spoken cov
 
 **Show:** README architecture, test output, and Cloudflare dashboard or deployment URL.
 
-**Say:** “GPT‑5.6 and Codex helped design the product, map multiple OpenStax sections into simulations, implement the Cloudflare Worker and D1 schema, write the scientific and end-to-end tests, perform a harsh product audit, and iterate the responsive interface. The deployed runtime remains Cloudflare Workers, Workers AI, D1, and static assets.”
+**Say:** “GPT‑5.6 through Codex generates structured lessons and chooses only from nine reviewed simulation runtimes; NVIDIA generates bounded study questions and note answers. Cloudflare Workers and D1 keep every key and learner record server-side, and deterministic TypeScript—not generated JavaScript—produces the scientific evidence.”
 
 End on the product name and one sentence: **“Perception Learning makes reasoning visible while it can still change.”**

DESIGN.md

@@ -11,8 +11,8 @@
 
 - Primary direction: Living field guide — an editorial library with paper, ink, subject-colored book cloth, margin notes, and laboratory instruments.
 - Secondary influences: modern children’s science museums, beautifully indexed reference books, and restrained game HUDs.
-- One memorable visual moment: a textbook section opens into a split-page “proof desk,” while the account page reads like a dark field-guide cover opening onto a precise paper sign-in form.
-- Explicitly avoid: dark AI dashboards, purple gradients, generic card grids, excessive pills, fake chat interfaces, reward confetti, and decorative gamification without learning value.
+- One memorable visual moment: a textbook section opens into a split-page “proof desk,” while Moss, the hand-painted seed-and-leaf guide shared with Environments, listens and speaks from the page margin.
+- Explicitly avoid: dark AI dashboards, purple gradients, generic card grids, excessive pills, full-page generic chat interfaces, reward confetti, and decorative gamification without learning value.
 
 ## Layout System
 
@@ -70,18 +70,20 @@
 - Buttons: rectangular, decisive labels, visible icon only when meaningful.
 - Inputs: paper-white, strong labels, 44px minimum touch targets.
 - Book cards: physical spine/cover composition with subject color and progress bookmark.
-- Navigation: library, lab, notes, trail, and a dedicated account page; mobile bottom bar keeps learning destinations separate from account controls.
+- Navigation: library, lab, notes, lesson creation, and a dedicated account page; mobile bottom bar keeps learning destinations separate from account controls.
 - Account page: editorial “learning passport” composition, with a dark field-guide story panel and a quiet paper form; never use a floating auth modal.
 - Empty states: suggest one concrete next action and one curriculum example.
 - Errors: plain-language recovery beside the failed tool.
 - Loading: model-building status with progress checkpoints, never a fake AI chat stream.
+- Learning companion: Moss is a persistent guide, not a separate destination. Its translucent conversation panel uses flat assistant turns, quiet learner bubbles, a compact voice-aware composer, and page-reading controls while remaining subordinate to the lesson.
 
 ## Accessibility Rules
 
 - Minimum body size: 16px; teacher metadata may be 12–14px.
 - Contrast: WCAG AA minimum.
 - Focus style: 3px translucent ultramarine outline with 3px offset.
 - Keyboard navigation: every book, reader section, note action, account control, and lab challenge is reachable.
+- Companion: launcher, quick prompts, narration transport, voice selection, and composer are keyboard reachable; responses use a polite live region.
 - 3D lab: equivalent text state through `render_game_to_text`; no visual-only success condition.
 
 ## Implementation Notes
@@ -92,3 +94,4 @@
 - Persistence: InsForge owns authentication; the authenticated user UUID scopes Cloudflare D1 notes, experiments, assessments, lesson guides, and progress so learning restores with the account.
 - Curriculum content: link and attribute OpenStax. Do not ingest entire books into Workers AI; use teacher-reviewed briefs and authored summaries.
 - Targets: current evergreen browsers, iPad/tablet landscape, 390px mobile minimum.
+- Narration: rank installed premium, enhanced, neural, and natural voices before generic defaults; preserve text as the fallback when speech synthesis is unavailable.