fix: hide diff scope labels for visible declarations

The visible declaration already provides scope context. Cache the actual
identifier point separately from the declaration extent/decorators and hide
labels when it is visible in that side of the projected viewport.

Refill reclaimed rows and anchor carets against the resulting layout.
Retain sticky context when the symbol scrolls or wraps out of view.

Verified: 951 tests passed, cargo build, and real PTY smoke for declaration
visibility/scroll restoration, wrapping, horizontal clipping, unified view,
and short horizontal-pane caret reflow.

docs/editing.md

@@ -438,8 +438,8 @@ Review marks belong only to the open comparison and survive filtering, file navi
 
 Text panes show change excerpts with **10 unchanged lines above and below** each change, preserving original file line numbers. Overlapping or touching excerpts merge; omitted stretches show an unchanged-line count in both panes. Files with only metadata changes show their notice or “No text changes” rather than unrelated source.
 
-**Scope labels** identify named enclosing functions/classes and containers, such as `impl Review › run`. Muted, unnumbered rows appear at excerpt starts and scope changes; the current scope stays at the top when its declaration scrolls out of view. Each split side uses its own source, so renamed functions keep distinct old/new names. Unified labels use `−` for Before, `+` for After, and `±` for shared names. Narrow panes prefer the innermost name. Labels are presentation only: motions, selections, copying, editing, and line jumps retain original source coordinates. Context comes from the existing Tree-sitter trees, including supported injected code, without requiring an LSP; unsupported syntax gets no guessed label.
-Within an excerpt, additional hunks and alignment gaps do not repeat an unchanged scope label. A new excerpt or an actual scope change gets a fresh label; sticky context is retained independently for each side.
+**Scope labels** identify named enclosing functions/classes and containers, such as `impl Review › run`. Muted, unnumbered rows supply context only when the declaration's symbol is outside that pane's viewport. Scrolling the symbol back into view removes the label and reclaims its row; decorators alone do not count as the declaration name. Visibility follows source columns and wrapped segments independently for each side. Each split side uses its own source, so renamed functions keep distinct old/new names. Unified labels use `−` for Before, `+` for After, and `±` for shared names. Narrow panes prefer the innermost name. Labels are presentation only: motions, selections, copying, editing, and line jumps retain original source coordinates. Context comes from the existing Tree-sitter trees, including supported injected code, without requiring an LSP; unsupported syntax gets no guessed label.
+Within an excerpt, additional hunks and alignment gaps do not repeat an unchanged scope label. A new excerpt or an actual scope change gets a fresh label when needed; sticky context is retained independently for each side.
 
 In **Normal mode** with a text pane focused, **`[` expands above** and **`]` expands below** the excerpt containing the caret, revealing up to **15 lines** per keypress in either pane. In editable After's Insert mode these keys insert literal text. Expansion clamps at file boundaries and merges excerpts when they meet. Existing visible source stays anchored. Opening a file resets its excerpts to the configured initial context. Live edits recompute changes while retaining expanded context; the caret stays visible even when the last text change is reverted.