feat: switch Qyburn to unified diff output + patch apply

Full-file replacement confirmed failing at all model sizes (tested up to
codestral:22b on 619-line file). Model hallucinates a new short file instead
of reproducing the original with changes applied.

New approach: model outputs only the unified diff (10-20 lines), applied
programmatically via `patch`. Works at any file size, uses ~154 tokens vs
~700+ previously, and succeeds on first attempt.

- Add `patch` binary to Dockerfile (apt install)
- Replace `_extract_code` with `_extract_diff`
- Add `_apply_diff` using subprocess patch
- Rewrite `_CODER_SYSTEM` to ask for diff format
- Rewrite `_build_prompt` to return plain string (no prefill dict)
- Rewrite `_ask_coder` to drop prefill assistant message
- Reduce num_predict 16384 → 4096 (diff output is short)
- Update `run_coding_loop` inner loop to apply diff then syntax-check result
- Update `_write_handoff` to show last diff instead of last full-file attempt

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.gitignore

@@ -0,0 +1,4 @@
+sandbox/
+jobs.json
+*.bak
+*.bak-*