E2E strategy: browser-level coverage for the walkthrough and builder flows

[adamstankiewicz]

Unit and integration coverage now guard the pure logic, the protocol envelope, and the format surfaces (stack #93). The remaining tier is browser-level e2e: the student walkthrough (advance, review, remediation injection, star economy), the builder's streaming UI, and the widget shell inside a host frame.

Prerequisites worth deciding before writing any specs:

• Model-call stubbing: e2e must run offline — either a fixture-replay mode for streamPathway/buildWidget (record NDJSON streams as fixtures, replay behind an env flag) or a fake provider in model.ts. Fixture-replay preserves the real route code paths and is the better instrument.
• Tooling: Playwright (already proven against the shell manually — the feat: self-building widget shell with single-source theming #86 functional verification was a Playwright pass).
• Scope discipline: e2e guards flows, not pixels — walkthrough completion end-to-end with evidence persisted, remediation injecting after a wrong finish, the shell rendering a spec inside an iframe with host-bridge messages. Visual regression is out of scope.

Sequencing: after stack #93 merges; pairs naturally with the Claude Desktop QA plan (the manual passes it can't automate stay manual).

🤖 Generated with Claude Code