feat(playwright): unwrap report.json suite + reasoning-driven screen capsule

[drewstone]

Why

run-capsule --playwright report.json silently produced an empty video for @tangle-network/browser-agent-driver (bad) runs. Two gaps:

1. Suite wrapper. spansFromPlaywrightResult read .agentResult / .turns, but the driver's canonical on-disk artifact is a TestSuiteResult ({ results: TestResult[] }). Neither key existed at the top level → turns = [] → 0 spans → blank capsule.
2. Reasoning dropped. The driver records per-turn agent reasoning on attributes.reasoning, but the screen capsule only captioned the action + URL — so a browser run never rendered as the "thinking-driven" screencast it could be.

What

• Adapter (adapters/playwright.ts): accept the suite shape and map results[0] (first case by default; multi-case selection stays a CLI concern). Bare Turn[], single TestResult, and agentResult inputs are unchanged.
• Screen capsule (renderers/screen-capsule.ts): thread reasoning from attributes.reasoning (fallbacks: attributes.thought, result.reasoning) through ScreenStep into the caption as a 💭 line (whitespace-collapsed, clamped to 280 chars + 2 CSS lines).

Verification

End-to-end against real bad report.json files (not fixtures):

• 2-turn run → screen.mp4 (1280×720, 6.3s) + replay.mp4 — real product frames, action captions.
• 24-turn Variant run → screen capsule with the 💭 reasoning line over each real frame (e.g. "Fixed the script syntax by wrapping it in an IIFE. This will properly set the input value…"), step "7/24".

Gate: pnpm typecheck clean, pnpm test 43/43 pass (2 new: suite-unwrap, reasoning-on-step).

Scope

Additive and backward-compatible — no field removed or renamed; every previously-accepted input shape still maps identically.