Recommend Role Skills

Layer-2 recommender: suggest which specialist skill rounds this role's JD demands (weight 1-3 + a JD-grounded rationale). A PROPOSAL the recruiter approves/tunes — not a decision. Behavioral is always-on and never listed.

RUNNABLE-ness is NOT filtered here. The recommender matches against the full skill LIBRARY (specialists/library.py), which deliberately carries signatures for skills whose assessment module doesn't exist yet — so this route can return e.g. analytical_data while nothing can actually run it.

Three guards downstream keep that from reaching a candidate, and they are where the filtering genuinely lives:

• GET /api/interviews/available-skills returns only runnable entries, so the schedule form and the assessment-plan editor never offer the rest
• the plan editor filters suggestions against that same list again
• the brain's orchestrator raises "Unknown skill: <id>" if a session is somehow started for a skill with no module

Do not "fix" this by filtering here: an unrunnable suggestion is the signal that a customer's JD wants a skill we haven't built, which is exactly the input for deciding what to build next.