Codify browser-first foundational teaching in the PRD

[Doppp]

Context

PRs #14 and #15 implemented and hardened the self-contained AI foundations course. This PR closes the source-of-truth documentation gap without changing the runtime, schema, directives, lesson IDs, or progress model.

Milestones

1. Define canonical lesson content: foundational definitions, notation, worked explanations, and exercise context live in browser-rendered Markdown; tutor chat adapts and clarifies them.
2. Align normative examples: add the orientation stage, update tutor policy, replace the skeletal authoring example, and add foundational-content acceptance criteria.
3. Record the release: update the PRD to Draft v0.4, changelog, implementation status, and verification record.

Each milestone is represented by a focused commit. No implementation-plan file is added or changed.

Teaching contract

• define the terms and notation required for a meaningful prediction before asking it
• delay the full mechanism until the learner has evidence when discovery supports that sequence
• connect observations to a worked formal explanation, code or data shapes, a deliberate failure, and an explanation-based recap
• keep instructional depth under editorial review and learner playtesting rather than word-count validation
• use Codex as the adaptive tutor, not the only source of foundational concepts

Validation

From a clean detached worktree:

• pnpm check: 33 suites, 102 tests
• pnpm course:test: 17 suites, 54 tests, plus all 14 starter/reference pairs
• pnpm build
• pnpm test:browser: 16 tests including axe and 320px coverage

The required five-learner study remains external acceptance evidence and is not represented as complete.