Stage 1 pre-filter for large skill catalogs: when an agent has more unpinned skills than
maxSkillDecisionCandidates (default DEFAULT_MAX_SKILL_DECISION_CANDIDATES), ranks
the unpinned skills using RankSkillsByHybridRelevance and retains only the top M
for Stage 2 (AgentDecisionService), pre-hiding the tail.
Stage 1 pre-filter for large skill catalogs: when an agent has more unpinned skills than
maxSkillDecisionCandidates(default DEFAULT_MAX_SKILL_DECISION_CANDIDATES), ranks the unpinned skills using RankSkillsByHybridRelevance and retains only the top M for Stage 2 (AgentDecisionService), pre-hiding the tail.