Implement planner-executor-critic scoring pipeline

Problem
Current scoring is a one-shot model step that outputs score schema, missing-signal data, and recommendations in a single pass. This allows malformed JSON and logical inconsistencies to pass through unchecked.

Impact
Unstable local model outputs reduce trust, can break downstream consumers, and make missing-signal interpretation unreliable. Contradictions such as mismatched missing keywords and gap analysis can silently appear in scored outputs.

Proposed Fix
Introduce a 3-stage local loop: Stage 1 Planner proposes the score schema and missing-signal hypotheses. Stage 2 Executor generates score and recommendations based on the planner output. Stage 3 Critic validates JSON shape and contradiction rules (for example, skills listed in missing_keywords must exist in gap_analysis and vice versa), and can request one corrective retry when validation fails.

Functionality Impact
Scoring becomes more deterministic and self-correcting, with more stable and trustworthy outputs. The loop adds an extra model/validation cycle and slight latency increase, but should reduce malformed payloads and improve downstream reliability.