v2: Full-page capture beyond the viewport

[loyality7]

Context

v1's screenshot only captures the visible viewport. Two candidate mechanisms exist and the choice needs a spike, same rigor as v1's protocol-backend spike:

1. The process-global whole-document draw setting — must be applied before any view exists, so it becomes a library initialisation option, not a per-call flag.
2. Scroll-and-stitch.

Scope

• A spike measuring both approaches: correctness on a long/complex page, memory cost, and time cost.
• Written GO/NO-GO per approach, then implement whichever wins.

Acceptance criteria

• Spike produces measurements for both approaches on a real device
• Written decision recorded before implementation starts (same discipline as the v1 protocol spike)
• Chosen approach captures a page taller than the viewport correctly, verified against a fixture page built for this