feat: achieve 100% ground truth coverage (248/248 tasks)

Add three new extraction strategies to ground_truth.py:
- MCP oracle extractor: reads tests/oracle_answer.json for all 81
  ccb_mcp_* tasks (files/symbols/chain arrays → evidence_files)
- instance.json check in _gt_sdlc: covers 7 deps-install tasks
- Schema F/H handlers: registers 10 doc-gen tasks as gt_type=generate

Also fixes:
- Case-sensitivity bug in _resolve_task_dir (CCX-* → ccx-* folding)
  that silently dropped 61 MCP tasks
- Upgrades 20 existing MCP entries from garbage test_script fallback
  to high-confidence oracle data
- Adds hardcoded GT for 8 remaining SDLC tasks (django-*, test-*)
- Adds __main__ CLI for standalone registry regeneration

Coverage: 165/248 (66%) → 248/248 (100%)

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

Author: sjarmak