fix(eval): enforce deciding evidence in knowledge promotion

[drewstone]

Memory activation and RAG promotion now require Eval's complete deciding result.
Insufficient or indeterminate evidence holds promotion, including tied binary safety results whose diagnostic interval is zero-width.
Memory runs capture source-unit mappings before execution and bind them to resume identity.

Adopt Eval 0.181.0 for Knowledge 17.0.0, migrate the held-out release decision type, and derive package checks from the declared peer range.
Document the existing claim and durable final-evidence controls, with an integration test through the actual retrieval wrapper.

Validation:

• Native Linux Node 22 full suite: 883 passed, 0 failed, 12 optional integration tests skipped; typecheck and build passed.
• Lint, API surface, skills, containment, and version-bump checks passed.
• Packed-package verification passed against registry Eval 0.181.0, including clean install, declarations, imports, CLI, publint, and attw.
• Official GEPA and SkillOpt packed integration: 2 passed with agent-eval-rpc 0.181.0.
• Source archive SHA-256: 4882bd58187261fd6e1a5880d677ddfac0408e072fc43e6a78cee9c008c2ebe3.

The native run excludes macOS metadata files accidentally included in an earlier transfer.
The package's Linux candidate-containment guard remains intact; optional provider integrations need their external services.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.