Single-stage XML review: eliminate extraction call, inline <issue> tags

- Refactor to single agentic loop that outputs <issue> XML tags directly
- Remove separate extraction API call (was a second LLM call per review)
- Accumulate text across all agentic rounds to capture mid-investigation findings
- Robust XML parser handles edge cases (multiline, missing fields, greedy comments)
- Rename warpgrep_codebase_search -> codebase_search
- Increase WarpGrep searches from 4-6 to 8 minimum
- Add spelling check, callback verification, nil-check patterns to prompt
- Tighter dedup (0.25 same-cat, 0.40 cross-cat) and cap at 4 issues/PR
- Results: R=65.1% (was 48.8%), P=68.3%, F1=66.7% (was 62.7%)