Add entity-level (semantic) diff understanding to the Code Review panel,
so diffs are presented in terms of code entities (functions, classes,
structs) rather than raw lines.

What:
- New semantic_diff module in crates/languages with entity extraction,
  3-phase matching (exact name → structural hash → fuzzy similarity),
  and change classification (modified/renamed/added/deleted/moved/unchanged)
- SemanticDiff feature flag (dogfood-enabled) gating all new behavior
- Entity summary pills in file sidebar (e.g. 'fn foo (modified) ·
  bar → baz (renamed) · struct X (added)')
- Entity-aware stats in file header (e.g. '+5 -3 (2 modified, 1 renamed)')
- Entity diff computed at FileDiffAndContent construction from base and
  current file content using existing tree-sitter infrastructure

Why:
- Line-based diffs with ±4 context don't align with how humans read code
- Renamed functions show as +5 -5 rather than 'renamed'
- Formatting-only changes are indistinguishable from logic changes
- Upstream issue warpdotdev#9898 requests this capability

How:
- Reuses existing tree-sitter parsing (arborium) and identifiers.scm
  queries — no new parser dependencies
- Body hash skips signature line so renames produce matching hashes
- Whitespace-normalized hashing so formatting-only changes classify
  as Unchanged
- Falls back to standard line-based diff for unsupported languages
  (no identifiers.scm) or parse failures
- Works on all platforms including WASM (entity_diff is None when
  local_fs is unavailable)

Tests: 19 unit tests covering extraction, matching, rename detection,
formatting-only classification, and multi-language support (Rust, Python,
JavaScript). All existing code_review tests (75) and diff_state tests (20)
continue to pass.

Co-Authored-By: Warp <[email protected]>

Fix 6 issues from review on PR warpdotdev#9917:

1. fix(code_review_view): update entity_diff on watcher refresh
   - Add current.entity_diff = diff.entity_diff.clone() alongside
     current.file_diff = diff.file_diff in update_from_single_file_diff_result
     so entity summary doesn't go stale after edits

2. fix(diff_state): guard entity diff computation with diff-size limits
   - Skip entity diff for DiffSize::Large/Unrenderable files
   - Cap current file content at 4.375 MB before parsing
   - Prevents unbounded CPU/memory on large repo-controlled files

3. fix(diff_state): gate fs::read_to_string behind local_fs cfg
   - Wrap entity diff computation in #[cfg(feature = "local_fs")]
     with None fallback for non-local-fs/WASM builds
   - Both construction sites (diff_state_against_head and
     file_diff_for_path) are now properly gated

4. fix(semantic_diff): accept bare @Definition captures
   - Previous filter skipped bare @Definition (used by 18+ languages
     including C, C++, C#, Java, JavaScript) and leaked @reference.call
   - Now accepts any capture starting with "definition", rejects
     everything else
   - type_prefix extracted via strip_prefix("definition.")

5. fix(semantic_diff): mask name token instead of skipping signature
   - Skipping the first line made single-line entities hash the empty
     string (fn foo() { 1 } and fn foo() { 2 } classified as Unchanged)
   - Now masks (replaces with spaces) only the name token's byte range
     in the entity body, preserving signature structure while making
     the hash invariant to renames
   - Adds tests for rename-invariant hashing and single-line entity
     body changes

6. fix(semantic_diff): use neighbor-based Moved classification
   - Absolute line-shift heuristic caused false Moved when lines were
     inserted above an entity
   - Now compares immediate neighbors in the entity ordering: both
     neighbors must change for Moved, preserving robustness to
     insertions/deletions above
   - Adds test_insertion_above_does_not_cause_moved

Co-Authored-By: Warp <[email protected]>

…diffing

Add SemanticDiff feature flag (Dogfood) that enables two complementary
improvements to the code review diff panel:

1. Entity classification (crates/languages/src/semantic_diff.rs)
   - Extract semantic entities (functions, methods, structs, impls) from
     tree-sitter using existing identifiers.scm queries.
   - 3-phase matching across base/current: exact name → structural body hash
     → fuzzy Jaccard similarity (>80%).
   - Classifies changes as: Unchanged, Modified, Renamed{old_name}, Added,
     Deleted, Moved (via neighbor position shift detection).
   - Body hashes use name-token masking so renames alone don't affect the hash.
   - Gated behind #[cfg(feature = "local_fs")] with 4.375 MB file size cap
     and DiffSize::Large/Unrenderable guard.

2. Syntax-token diff (replacing line-based hunks for supported languages)
   - Parse both files with tree-sitter, collect leaf tokens (DFS).
   - Comments and strings treated as opaque (atomic tokens) to avoid
     fragmented partial matches within rewritten comment blocks.
   - LCS diff on token hashes via similar::TextDiff (same Patience algorithm
     as line diff, but on tokens).
   - Map token-level ops to line-level SyntaxHunks.
   - Multi-line rewrites → clean Deletion + Addition blocks (no inline noise).
   - Single-line changes → Replacement with per-word inline highlights
     (e.g. renamed function names).
   - Falls back to line-based diff for unsupported languages or parse failures.

3. Entity-scoped context (app/src/code/editor/model.rs)
   - calculate_hidden_lines() expands visible ranges to cover entire
     entities containing changes instead of fixed ±4 context lines.
   - Added EntityRange { start_line, end_line, kind } to CodeEditorModel,
     wired from diff_state entity_diff data.

4. removed visual noise per user feedback
   - Removed sidebar entity summary pills (render_entity_summary).
   - Removed file header entity counts (render_file_stats_with_entities
     reverted to render_file_stats).
   - These were duplicating existing diff information and adding clutter.

5. Infrastructure
   - Added file_path to DiffModel for language detection in syntax diff.
   - Added similar.workspace = true to crates/languages/Cargo.toml.
   - Added 30 unit tests covering entity extraction, matching, classification,
     syntax-token diff, word-level highlight alignment, and comment opacity.

Refs warpdotdev#9898