docs: specialize org governance for OCR evidence and model lifecycle

[e54-bot]

Summary

• AGENTS.md applies the organization testing, verification, engineering-quality, and entropy policies explicitly, keeping OCRKit's stricter recognition/model specialization local.
• Expected truth: recognition expected values require authority independent from the model/parser under test (reviewed labels, reproducible screenshot content, accepted layout/API contract, provenance-tracked regression); stable model output is not its own oracle and labels are not rewritten to make a new model pass; aggregate accuracy and fixture/test counts are evidence summaries; model versions, layout/fixture counts, and dataset cardinality are not durable contracts.
• Model/evaluation boundary: training evidence and the decisive held-out evaluation set stay disjoint (the versioned source-level split in training/README.md is the enforcement mechanism); mutable release pointers/channels are never fixture truth.
• Mechanism admission: fallback models, routing layers, and multimodal paths require the same measured recognition gap as training a new model.
• Entropy: OCRKit-specific cleanup targets (duplicate preprocess paths, obsolete model fallbacks, stale layout compatibility, duplicated parser normalization, ownerless generated artifacts, migration adapters) with a consumer-mapping requirement before removing recognition paths; uncertainty/status reporting, validation, privacy boundaries, artifact verification, and relied-on compatibility stay.

Test plan

• Docs-only change; each Specialize org governance for OCR evidence, evaluation, and model lifecycle #27 acceptance criterion verified against the final text
• git diff --cached --check clean; no code paths touched
• Linked organization policy documents confirmed present on OWBastion/.github main

Fixes #27