ReconciliationModule orchestration (numpy-native; drop wandb + multiprocessing)

[Polichinel]

Epic: #31 · Depends on: #34, #35

Background

Assemble the adapters (#33), the grouping core (#34), and the validation (#35) into the public reconciliation entry point — the ReconciliationModule equivalent. The original drags ProcessPoolExecutor (for torch/GPU dispatch) and wandb alerting; neither is justified for fast numpy arithmetic.

Work

• A single orchestration class/function: validate → group+reconcile → return a new pgm PredictionFrame.
• SRP: orchestration only — the math is the leaf, the mapping is injected, the I/O is the adapters, the validation is helpers. One main class per file; composition over inheritance.
• Drop wandb and ProcessPoolExecutor; document why (numpy is fast, no GPU). If scale ever demands parallelism, add it behind the same interface (OCP) — do not bake it in now.
• De-mutated return (C-184).

Acceptance criteria

• reconcile(...) runs end-to-end on the S0 fixture and returns a reconciled pgm frame.
• No torch / pandas / viewser / wandb imports anywhere in views_postprocessing/reconciliation/.
• Input frames are untouched.

Validation

• Integration test running the full module on the S0 fixture (precursor to the end-to-end parity gate, the next story).