Improve agent discovery and text expansion for large text surfaces

[thymikee]

Summary

Large text surfaces such as code editors, document views, chats, logs, and rich text panes are currently awkward for agent use.

Today snapshot -i may expose these surfaces by dumping a partial text blob, which is expensive in tokens and still incomplete for discovery. At the same time, get text is snapshot-derived, so if the snapshot text is truncated there is no reliable way to expand that same surface into the full visible text.

A concrete example is Android Studio on macOS: the editor is exposed as a TextView, but the interactive snapshot shows only a partial code fragment, while important visible content may still be missing.

Problem

We need a cleaner split between:

• discovery: what visible text surfaces exist and which one should the agent inspect next
• extraction: retrieving the text for a chosen surface after discovery

Without that split, snapshots either:

• spend too many tokens on giant text blobs, or
• truncate text in a way that makes the hidden content unreachable

Goals

• Keep large visible text surfaces discoverable in snapshot -i
• Make snapshot -i summarize those surfaces semantically instead of dumping long text bodies
• Add a reliable way to expand a selected text surface after discovery
• Keep the design transferable across macOS, iOS, and Android

Proposed Direction

1. Add runner-backed text extraction for element-targeted reads instead of relying only on the stored snapshot node.
2. Update snapshot -i rendering for large text surfaces (TextView, TextField, editor-like panes, etc.) to prefer semantic labels plus a short preview.
3. Mark truncation explicitly so agents know more content exists.
4. Include useful metadata when available, such as editable, scrollable, focused, or similar state.

Example desired shape:

@e32 [text-view] "Editor for MainActivity.kt" [editable] [scrollable] [preview:"package com.example..."] [truncated]

Acceptance Criteria

• snapshot -i shows large visible text surfaces as first-class nodes without dumping the full body by default.
• Agents can follow up with an element-targeted text read and retrieve the visible or full text for that surface.
• Truncation is explicit in discovery output.
• Behavior works consistently enough to support desktop editors on macOS and similar large text surfaces on iOS/Android.

Non-Goals

• OCR fallback for now. Agents can use screenshots separately when needed.

[thymikee]

Closed in #261