Qualify production model behavior and deployment identity for runnable agents

The v2 implementation is in PR#6: actual MetaHarness kernel, native RuVector scoped lexical memory, optional actual RuFlo keyword routing, bounded provider/tool loop, CLI/MCP, standalone project export with matching lockfile, README/ADRs/security/evidence and CI delivery.

Completed validation:23 tests; independent security review with ancestor symlink and bounded-read corrections; zero reported vulnerabilities in supported Linux x64 Node24 install; clean standalone npm ci, audit, CLI, fixture loop and replay.64 deterministic lexical queries achieved recall@3=1 and cosine-oracle agreement=1; context reduction94.15%, not model-quality evidence.

Operational work remains intentionally open:

• Qualify selected live model on representative tasks, adversarial cases and actual cost budgets before production use.
• Configure dedicated operator provider key/model/live-call permission and private durable storage. Never put credentials in manifests.
• Verify target hardware/native package support beyond qualified Linux x64 Node24.
• Provision authenticated deployment identity and storage for remote field memory if required; local receipts remain unsigned integrity evidence.
• Federation catalog tools do not execute in this agent runtime. Any future write adapter needs separately reviewed membership, signature and operator authorization enforcement.

No deployment, public federation message, automatic promotion or SOTA certification is claimed.

Start by reviewing PR#6 and the README, ADRs, security evidence, and CI delivery described in the issue. Define qualification and deployment checks for the live model, provider permissions, target hardware, deployment identity, durable storage, and any federation write adapter; done means these boundaries and authorization requirements are verified without claiming deployment or certification.