enhancement: auto-chunking for large diffs

[ajianaz]

Description

Full codebase review (500K+ chars) exceeds default max_diff_size (50K). Users must manually split diffs into batches.

Code References

File	Line	Current State
src/commands/review.rs	88-113	Size check — max_diff_size enforced, no chunking
src/engine/llm.rs	261-270	review_diff() — single LLM call for entire diff
src/config/schema.rs	default	max_diff_size: 50_000 (50KB)

No chunking logic exists. If diff > max → error or blocked (depends on on_violation setting).

Fix Plan

Architecture

review.rs
  ├── diff > max_diff_size && auto_chunk enabled?
  │   ├── chunk_diff(diff, max_size) → Vec<DiffChunk>
  │   │   ├── Parse diff into per-file hunks
  │   │   ├── Group by directory (related files together)
  │   │   └── Each chunk ≤ max_diff_size
  │   ├── for chunk in chunks:
  │   │   ├── review_diff_with_cache(chunk)
  │   │   └── collect findings
  │   └── merge + deduplicate findings
  └── else: single review (current behavior)

Implementation Steps

1. Add --auto-chunk flag to cora review CLI
2. New module src/engine/chunker.rs:
   • Parse unified diff into (filename, hunk_lines) tuples
   • Group by directory prefix
   • Split groups until each ≤ max_diff_size
   • Return Vec<String> (each a valid unified diff)
3. Update execute_review():
   • If --auto-chunk and diff > max → call chunker
   • Review each chunk separately
   • Merge ReviewIssue vectors
   • Deduplicate by (file, line, title) tuple
4. Progress reporting: "Reviewing chunk 3/10 (src/config/*)…"
5. Summary merge: concatenate chunk summaries

Edge Cases

• Single file > max_diff_size → warn, review truncated
• Overlapping findings across chunks → dedup
• Empty chunk → skip
• LLM failure on one chunk → partial results + warning

Target

• Minor release: v0.5.0 (new feature + new module)
• Files: new src/engine/chunker.rs, review.rs, main.rs, progress.rs

[ajianaz]

✅ Already implemented. Auto-chunking is enabled by default with --no-auto-chunk flag to disable. Uses src/engine/chunker.rs + execute_chunked_review() in review.rs.