feat: gold standard reference implementations for scenarios

[zenprocess]

Priority: High

PawBench has no ground truth outputs. SWE-bench has human-verified patches.

Proposal

For each PawStyle scenario, provide a reference implementation:

• scenarios/pawstyle-independent/reference/server.py — working backend
• scenarios/pawstyle-independent/reference/index.html — working frontend
• scenarios/pawstyle-independent/endpoints.json — expected API contract

Score generated code against the reference:

• API endpoints match (paths, methods, status codes)
• JSON response schemas match (field names, types)
• AST-level comparison for key patterns (not string matching)

Depends on: sandbox eval (execution-based scoring)