Summary

• Adds a full-screen telescope-style fuzzy search across every line of every file in the diff, opened with ctrl+f. Existing in-file / search is left untouched.
• Two-pane layout (50/50): results list on the left, real stacked-diff preview on the right with tree-sitter syntax highlighting.
• Powered by nucleo-matcher (same fuzzy engine as Helix / Zellij) for fzf-quality scoring.

What's in scope

Indexing (built once per modal open):

• Equal rows → new side only (no duplicates)
• Delete / Insert → only the side with content
• Modified → both sides (genuinely different text)
• Blank/whitespace-only lines skipped
• Smart case + smart unicode normalization

Results list:

• ❯ selector, change-symbol prefix, project-relative path:line, body with fzf-matched chars bolded
• Horizontal + vertical mouse scroll; list scroll separate from selection (mouse can scroll past cursor)
• Click a row to select; ↑↓ / ctrl+n / ctrl+p / ctrl+j / ctrl+k also select

Preview pane:

• Real stacked (unified) mini-diff around the matched line, full file scrollable
• Tree-sitter syntax highlighting, cached per (file, panel)
• Tab expansion matches main diff view
• Cursor row centered vertically; brightened in-palette so green-on-green / red-on-red are preserved, with selection_bg fallback on themes that defer to terminal bg
• Background colors match the main diff view (page bg for equal/empty, added_bg/deleted_bg for changes)
• Mouse wheel scrolls; shift+wheel flips to horizontal

Key bindings:

Performance

The render path is on the hot frame loop, so every layer is optimized:

• Two-pass preview render — pass 1 walks the full SBS building cheap (sbs_idx, side, is_cursor) metas; pass 2 materializes only the visible window through the expensive highlighter + tab-expand + scroll path. On a 1000-line file with a 30-row preview, that's ~33× fewer highlighter calls per frame.
• Run-coalesced spans in result rows — was one Span::styled(ch.to_string(), …) per haystack char (~4000 single-char allocs per frame for a typical visible list); now ~10× fewer spans by grouping consecutive same-style chars.
• Reusable Matcher — taken-and-restored across refilter calls so we don't pay nucleo's scratch-slab allocation per keystroke.
• Top-K via select_nth_unstable_by instead of full sort + truncate — ~10× faster sort phase on large match sets.
• SBS pre-seeded into cache during build — first preview render is cache-hot since we already compute SBS at index time.

Test plan

• cargo build and cargo build --release both clean
• ctrl+f opens; modal is full-screen with empty list + preview
• Typing a query updates results in real time; matched chars are bolded
• Preview pane renders a stacked mini-diff around the cursor row, syntax-highlighted
• Cursor row visibly highlighted on equal/added/deleted rows across themes
• enter jumps to the matched file + line and seeds / search
• Mouse wheel scrolls list and preview independently; shift+wheel horizontal
• Click a result row to select it
• cmd+backspace / ctrl+u clear query; opt+backspace / ctrl+w erase word
• esc / ctrl+c close without jumping
• Preview bg matches main diff view bg (no sticky-header tint)
• Smooth on a large diff (many files / many lines) — no visible lag on keystroke

🤖 Generated with Claude Code