Harden selective verification and complete competitor matrix

[samarth1412]

What changed

• Adds selective semantic_core_v2 and semantic_core_v2_1 verification, schemas, limits, concurrency isolation, and source-condition diagnosis.
• Adds controlled development corpora and engineering benchmarks for hard negatives, source conditions, and evidence attribution.
• Pins and audits RAGAS, MiniCheck, and RefChecker competitor runtimes.
• Completes the nine-row visible-development comparison matrix across RAGTruth, ARES, and the CRAG proxy track.
• Documents migration, research boundaries, and the SOTA engineering roadmap.

Why

The verifier needed safer selective behavior, stronger source-condition attribution, reproducible external comparisons, and explicit separation between visible development measurements and untouched evaluation evidence. The interrupted RefChecker run was caused by an Ollama auto-update from the pinned 0.32.1 runtime to 0.32.5; it was resumed and completed using the exact pinned runtime.

Impact

This gives ContextTrace a reproducible development baseline and stronger implementation safeguards without publishing a release, touching sealed labels, or claiming SOTA. CRAG remains an agreement/proxy track rather than an accuracy evaluation.

Validation

• 609 passed, 1 skipped.
• Coverage 86.85%; the 80% gate passed.
• ResourceWarning checks passed.
• All nine comparison rows reproduced from raw artifacts with exact ID, input, and hash binding.
• Targeted Ruff and mypy checks passed.
• git diff --check passed.