feat: build unified 280-task manifest organized by task type

Replace the suite-centric 220-task core manifest with a task-type-centric
unified manifest (schema v2.0). All tasks are peers — no SDLC/Org split.

Allocation: comprehension=100, implementation=90, quality=90 (280 total)
from a pool of 397 eligible tasks (test.sh + ground_truth.json required).

Key metrics:
- Overall statistical power: 84.1% at sigma=0.20, delta=0.05
- Large codebase (2M+ LOC): 58.6% of tasks
- Multi-repo: 31.8% (38% comprehension, 21% implementation, 36% quality)
- All 5 LOC bands covered in each task type
- 20 suites represented, 11 languages
- LOC fallback chain: selected_benchmark_tasks → promotion manifest →
  repo_set_id median → task.toml (eliminates all "unknown" LOC)

Selection algorithm: per-type round-robin across suites with LOC band
coverage, multi-repo representation, and core_ready preference.

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

Author: sjarmak