fix(ai): prioritize resolver semantics in capability discovery

[first-assist]

Summary

• rank resolver-authored descriptions and root operation identity ahead of incidental nested-field vocabulary
• tokenize PascalCase public operation/entity names and retain explicit result shape only as a semantic tie-breaker
• refresh package 0.10.4 release metadata, the coexistence lock, and overdue governed-document review checkpoints

Safety

• shape metadata alone still cannot admit an unrelated capability
• exact namespace, kind, entity/operation, authority, and execution checks are unchanged
• committed diff and PR text contain no deployment-specific vocabulary

Verification

• 37 tool-profile unit tests and 2 integration tests

• complete reverse-dependent AI default/SQLite suite (255 unit tests and all integration suites; 38 generated doctests intentionally ignored)

• warnings-denied Clippy and Rustdoc for the changed package

• locked root and coexistence metadata; package release, release manifest, workspace dependency, documentation governance, inventory, formatting, and SemVer gates

• Documentation updated

• No documentation impact