feat: add auditable agent task experiments

[SuTang-vain]

Summary

• harden extraction, asset resolution, Data Surface import, equivalence coverage, rules, and report path behavior
• add versioned Agent Task manifests, isolated patch execution, file-policy enforcement, task-specific grading, and runtime/visual evidence adapters
• add reproducible success-rate experiments with scripted and real-AI trial artifacts, digest audits, relocation checks, and explicit claim-level boundaries
• document task, patch, grader, evidence, TaskRun, and experiment contracts

Validation

• node --test tests/*.test.js: 188/188 passed
• git diff --check: passed
• credential-pattern scan: no findings
• artifact audit: 30 TaskRuns and 188 artifacts verified
• scripted experiment: 15/15 valid passes (harness-only)
• real-AI experiment: 7/14 valid passes, 1 infra error (ai-static-contract-only)

Scope note

The committed reports intentionally do not claim complete runtime/visual component-library capability where producer evidence is unavailable.