Add evidence-backed semantic skill ranking

[CountableNewt]

Goal

Add optional embedding-backed ranking only after the deterministic #24 resolver is deployed and measured.

Requirements

• Configure a real provider/model with readiness checks and explicit failure behavior.
• Use semantic ranking only as a supplement to required scope, explicit assignments, conflicts, enforcement, and capability rules.
• Define provider-neutral configuration and secret handling.
• Add offline relevance fixtures and compare deterministic-only versus hybrid precision/recall.
• Keep telemetry opt-in and redacted; do not retain full prompts, skill bodies, or source code.
• Expose configuration in the dashboard only after the backend consumes and verifies it.

[linear-code]

MCP-7