Implement 'one file at a time' diff generation strategy

MAJOR CHANGE: Generate diffs one file at a time instead of all at once

Strategy:
1. First API call: Identify which files need modification
2. For each file: Generate complete diff individually
3. Combine all file diffs into final patch
4. Hard validation before patch application

Benefits:
- LLMs handle single-file diffs much better than multi-file
- Each diff is guaranteed complete (has ---, +++, and @@)
- Eliminates truncated diff issues
- Falls back to single-pass if file identification fails

Additional improvements:
- Hard validation: Check for +++ b/ and @@ markers before applying
- Verify each file section is complete
- Better error messages with specific validation failures
- Improved prompts with stricter format requirements
- File existence verification

This should completely solve the '+++ b/p' truncation problem.

Author: htilly