fix: handle stalled and stale full-page captures

[iuyo5678]

Summary

Full-page capture needs an explicit stopping range when a site keeps its loading indicator visible. Some Windows browser builds also return stale window-surface pixels after programmatic scrolling. This change handles these separately while retaining the shared scrolling, tiling and PNG pipeline.

Capture range and diagnostics (all platforms)

• Add --scope current to capture the initial CSS document height. Default follow continues to follow appended content. The CLI refuses to save current unless the extension acknowledges the scope.
• Use the same bounded height calculation for begin, move and inspect, without changing the bottom-settling timers during begin.
• Report loading_stalled after 30 seconds at an unchanged loading bottom, with an actionable --scope current hint. Preserve distinct cancellation/interruption reasons and progress. Failed captures save no partial image.
• Keep both bundled skill copies, schemas and documentation in sync. The current scope is a height boundary, not a frozen DOM snapshot.

Windows Agent screenshot source

• Only Windows Agent captures prefer their existing session renderer. On macOS/Linux, a working captureVisibleTab source is retained; a failed initial probe falls back to the session renderer.
• Enable compact frame-freshness checks only for that Windows Agent path. Popup captures retain their existing source selection and do not run the new freshness check.
• Retry a clearly stale exposure twice before returning stale_frame. Layout mismatches and stale exposures have separate consecutive-failure counters, so mixed failures cannot produce the wrong recovery reason. Backends are selected before measuring and never switch mid-image.

Validation

• Extension: 1,682 unit tests passed; 94 opt-in tests skipped. TypeScript checking, lint and production build passed.
• Regression coverage includes Windows/macOS/Linux source selection, unchanged Popup behavior, both orders of mixed layout/stale failures, and height growth during begin.
• Direct source/engine verification in isolated Windows x86-64, headless Chrome for Testing 153.0.8010.36: after a 650 CSS-pixel scroll, native and renderer samples both matched the expected document row. The persistent-loader current capture produced a 1563 × 4800 PNG with zero mismatches in the sampled pixel column across all 4800 rows.
• Fault injection using a real frozen viewport PNG returned stale_frame after the initial accepted frame, before any repeated rows were written; the renderer-backed capture then completed correctly. This is controlled fault injection: the original spontaneous surface-cache fault did not reproduce on this Chrome build.
• Limits: macOS real-browser capture, the complete CLI/browser harness and the 125% scale renderer matrix remain unverified locally. Earlier full local Rust testing stopped on two existing Windows Hermes home-path tests; protocol and targeted CLI tests passed. These paths are outside this patch.