shep-kit: improve evidence collection to use main app and capture video for ui flows

[arielshad]

Problem

Currently, the evidence collection during implementation gathers evidence from Storybook. While this is useful for component-level verification, it has limitations:

1. Storybook-only evidence — screenshots come from isolated Storybook stories, not the actual running application. This misses integration behavior, real data flows, routing, and how components work together in context.
2. No video evidence for UI flows — for multi-step UI interactions (e.g., opening a drawer, submitting a form, navigating between views), static screenshots don't capture the full user experience. Video recordings would provide much stronger evidence.

Expected Behavior

1. Evidence from the main app

• Fine-tune the evidence collection prompt so it prioritizes capturing evidence from the actual running Next.js app (pnpm dev:web) rather than defaulting to Storybook
• Evidence should show the feature working in its real context — with real routing, layout, data fetching, and integration with surrounding components
• Storybook evidence can supplement but should not be the primary source

2. Video evidence for UI flows

• When the feature involves UI interactions (clicks, transitions, drawer open/close, form submissions, navigation), capture short video recordings as evidence
• Videos should demonstrate the complete user flow, not just the end state
• Use Playwright or similar tooling to record browser interactions programmatically

Acceptance Criteria

• Evidence collection prompt updated to target the main app for primary evidence
• Screenshots captured from the running Next.js app showing features in real context
• Video recordings captured for UI-related flows demonstrating multi-step interactions
• Storybook evidence retained as supplementary (component isolation) but not primary
• Evidence artifacts (screenshots + videos) stored alongside spec deliverables

[arielshad]

Most of this issue has been addressed through several merged PRs, but one gap remains:

Addressed

• App-level evidence over Storybook — PR feat(agents): require app-level proof in ui evidence validation #387 added validateUiEvidenceHasAppProof which detects Storybook-only screenshots and warns that app-level proof is required. Evidence prompts now mandate screenshots from the running dev server, with Storybook as supplementary only.
• Evidence validation + retry loop — PR feat(agents): add evidence validation and retry loop to evidence collection #419 added a retry loop (up to 3 attempts) that checks evidence quality and provides structured feedback to the agent when app-level screenshots or other evidence types are missing.
• Evidence framework — PR feat(agents): add pr evidence collection with mandatory screenshots #291 established the evidence collection pipeline with support for screenshots, test output, terminal recordings, and video types. PR feat(agents): add incremental commits to implement and evidence prompts #426 added incremental commit handling during evidence collection.

Remaining gap

• Proactive video capture for UI flows — The infrastructure supports video evidence (validation recognizes the video type), but there's no prompt or tooling that actively instructs the agent to record browser interactions for multi-step UI flows (e.g., using Playwright's video recording). The acceptance criterion "Video recordings captured for UI-related flows demonstrating multi-step interactions" is not yet fulfilled.

This issue could either be scoped down to just the video capture gap (and the other acceptance criteria checked off), or closed with a new issue tracking the video recording feature specifically.